The accuracy–fairness–efficiency Trilemma in mobile image classification: a Pareto benchmark Authors: Thanh TranVan, Hien LamThanh, Toan DoNang, Tuan-Tu Huynh Published: 2026 ABSTRACT AbstractDeploying deep learning classifiers on resource-constrained mobile devices requires simultaneously satisfying three competing objectives: predictive accuracy, demographic fairness, and inference efficiency. No prior work has jointly formalized these as a constrained multi-objective optimization problem, benchmarked a comprehensive strategy set under identical conditions, or defined a Deployment-Feasible Zone (DFZ) as the feasibility-constrained Pareto subset. This paper provides all three contributions. We benchmark eleven optimization configurations on a 2,821-image dataset with 24 intersectional demographic subgroups (worst-case imbalance 35.47:1) under hard constraints (, per attribute, MB, ms on an entry-level SoC). Three findings emerge. (1) The combination of 3D-aware augmentation and Protected Fairness Pruning (C2) is the Pareto knee point: (95 % CI: 0.906–0.962), , 6.3 MB, 187 ms, confirmed in 94.2 % of bootstrap resamples. (2) Standard magnitude pruning is strictly Pareto-dominated by fairness-constrained pruning (PFP) at identical compression ratio – a result grounded in the low-magnitude minority-encoding effect. (3) The Adaptive Trilemma Weight Scheduler (ATWS) yields consistent gains of +1.3 pp and -0.6–0.7 pp EOD over fixed-weight training, compatible with any fairness-constrained strategy.Keywords: Multi-objective optimization, Pareto frontier, Demographic fairness, Model compression, Knowledge distillation, Quantization, 3D data augmentation, Image classification, Edge deploymentSubject terms: Engineering, Mathematics and computing Introduction Deploying neural image classifiers on mobile devices requires the simultaneous satisfaction of three constraints that are individually well-studied but jointly under-theorized. First, a minimum predictive performance threshold must be met: a classifier with below 0.85 provides insufficient decision support regardless of its compactness1. Second, demographic fairness constraints must be respected: a model whose True Positive Rate differs by more than 10 percentage points across gender or ethnicity subgroups implements systematic discrimination independent of aggregate accuracy2,3. Third, device constraints impose hard limits on model storage size (łe 10 MB for over-the-air update compatibility) and inference latency (łe 300 ms for interactive use on entry-level SoCs)4,5. These three constraints induce a trilemma: optimization along one dimension typically degrades at least one other. Increasing model capacity improves accuracy but violates size constraints. Aggressive magnitude pruning reduces size but disproportionately removes neurons associated with minority subgroups6. Stronger fairness regularization can reduce accuracy below the minimum threshold7. Multi-objective optimization (MOO) has been applied to accuracy–efficiency trade-offs in NAS5,8–14 and to accuracy–fairness trade-offs7,15–17, but no prior study has jointly formalized all three objectives, benchmarked a comprehensive strategy set under identical conditions, or identified the deployment-feasible Pareto subset. This paper closes all three gaps with four contributions: (C1) Trilemma Formalization: We formalize the Accuracy–Fairness–Efficiency Trilemma as a constrained MOO problem with normalized, dimensionally consistent component losses and an explicit distinction between per-attribute and aggregate fairness2,15. (C2) Systematic Pareto Benchmark: We evaluate eleven optimization configurations under a fixed backbone (DenseNet-12118), fixed dataset partition, and leak-free augmentation protocol, reporting 95 % bootstrap confidence intervals and 5-fold cross-validation statistics. (C3) Deployment-Feasible Zone and Pareto Analysis: We introduce the DFZ as the feasibility-constrained Pareto subset19,20 and confirm C2 as the knee point in 94.2 % of bootstrap resamples across a ternary sensitivity analysis of 36 weight combinations21. (C4) Adaptive Trilemma Weight Scheduler (ATWS): We propose a phased-weighting algorithm that mitigates gradient conflict in multi-objective training. ATWS is an original contribution of this paper; the gradient-conflict phenomenon motivating it is documented in the multi-task optimization literature22,23, though those works do not address the three-objective fairness–accuracy–efficiency setting. Related work Fairness in machine learning Algorithmic fairness in supervised learning was formalized by Hardt et al.2 through the Equal Opportunity framework, establishing TPR parity across demographic groups as the canonical metric. Zafar et al.15 proposed convex relaxations for training-time fairness enforcement. Surveys by Mehrabi et al.3 and Caton and Haas24 document that most fairness research addresses a single protected attribute, while intersectional fairness — fairness over groups defined by multiple attributes jointly — exposes additional disparities invisible to single-attribute metrics. Kearns et al.25 demonstrated that optimizing fairness on individual attributes does not guarantee fairness for intersectional subgroups, motivating the 24-cell intersectional evaluation protocol used in this benchmark. Diana et al.26 proposed minimax group fairness (minimizing worst-group error) as a complementary formulation; the per-attribute EOD constraint in Eq. (9) approximates this within the feasibility-filter design. Rajkomar et al.27 extended these concerns to deployed AI systems, arguing that fairness-unaware models amplify existing health inequities. In the image classification setting, demographic disparities arise from dataset imbalance and model compression that preferentially degrades underrepresented subgroups. Efficient neural network deployment Post-training quantization28 reduces model size and latency by mapping float32 weights to INT8, typically incurring 0.5–2.0 pp accuracy loss. Mixed-precision variants29 enable finer-grained trade-offs. Knowledge distillation30 transfers capacity from a large teacher to a compact student, enabling MobileNetV34 and MobileNetV231 to approach larger-model accuracy. EfficientNet32 establishes a strong accuracy–efficiency baseline. Neural architecture search (NAS) under hardware constraints5,8–10 optimizes architecture itself but incurs computational overhead incompatible with limited-data settings; Karl et al.11 survey this landscape systematically. Critically, model compression is not fairness-neutral. Hooker et al.6 showed that compression amplifies bias concentrated in a subset of “Compression Identified Exemplars,” with magnitude pruning affecting 637 ImageNet classes significantly versus 119 for INT8 quantization. Stoychev and Gunes33 found that post-training quantization does not systematically increase gender or ethnicity gaps on facial expression datasets, consistent with the theoretical argument that quantization reduces precision uniformly while magnitude pruning removes neurons selectively in ways correlated with minority-group feature encoding. Dai and Li34 proposed bi-level optimization integrating fairness and pruning; Lin et al.35 demonstrated gradient-guided fairness-aware pruning (FairGRAPE). The present paper is the first to place representative strategies from all four major compression families on the same three-objective Pareto frontier under identical experimental conditions. Table 1 summarizes methodological differences among fairness-aware pruning approaches. Table 1Comparison of fairness-aware pruning approaches.PropertyPFP (ours)FairGRAPE35Dai & Li34Std. MagnitudeFairness criterionPer-attr EODGradient decomp.Bi-level obj.NonePruning granularityUnstructuredStructured (filter)UnstructuredUnstructuredCost (vs. std.)–5×–10×1×Differentiable fairness?No (discrete)YesYesN/AFixed-budget compatible?YesPartialNoYes Comparison of fairness-aware pruning approaches. Data augmentation for imbalanced demographic data Standard geometric augmentation applies transformations that distort structural attributes relevant to classification36. SMOTE37 generates synthetic minority samples via feature-space interpolation but does not respect image manifold structure. Fairness-aware augmentation using latent-space editing38 and 3D-aware synthesis address this by operating in semantically structured spaces. The DECA framework39 operating in the FLAME parametric space enables controlled variation of pose and expression while preserving the subject-specific shape parameter encoding classification-relevant structural attributes. Evaluating augmentation quality requires metrics that capture semantic fidelity rather than pixel-level similarity; Kebaili et al.36 and He et al.40 support using FID and LPIPS rather than SSIM/PSNR, since pixel-level metrics penalize geometrically valid variation. Multi-objective optimization in machine learning MOO seeks to characterize Pareto-optimal solution sets rather than a single optimum19,20. Chen et al.23 provide a taxonomy of gradient-based MOO for deep learning; MGDA41 and EPO Search42 are canonical two-objective methods. Özbulak et al.12 propose a multi-objective evaluation framework for utility–fairness trade-offs. Fairness-aware MOO has been studied by Liu and Vicente7, who formulate accuracy and fairness as independent MOO objectives. Candelieri et al.16 extend this to three objectives (accuracy, fairness, energy). Li and Li17 survey robustness–accuracy–fairness trade-offs, identifying the absence of a principled three-axis framework as an open problem. While Hu et al.43 prove theoretically that linear scalarization cannot recover all Pareto-optimal solutions on non-convex frontiers, the ternary sensitivity analysis (Section "Pareto frontier and sensitivity analysis") empirically verifies that C2 is the knee point in 66.7 % of weight scenarios, suggesting approximate convexity in the DFZ region20. Reiners et al.13 and Hotegni et al.14 apply MOO to sparsity–accuracy trade-offs in pruning and multi-task learning, but neither introduces a fairness dimension or deployment-feasibility filter; the PFP Pareto-dominance result is therefore new to both literatures. Comparative summary Table 2 summarizes representative prior work. No prior study simultaneously addresses all three trilemma axes with a formal Pareto framework and deployment-feasibility filtering. Table 2Summary of representative prior work. \checkmark: addressed; (\checkmark): partially; ×: not addressed.PaperTaskAccFairEffParetoDFZHardt et al.2Decision-making(\checkmark)\checkmark×××Zafar et al.15Classification(\checkmark)\checkmark×××Liu & Vicente7Classification\checkmark\checkmark×\checkmark×Candelieri et al.16HPO\checkmark\checkmark(\checkmark)(\checkmark)×Wu et al.5Mobile NAS\checkmark×\checkmark\checkmark×Karl et al.11HPO survey\checkmark×\checkmark\checkmark×Jacob et al.28Image classif.(\checkmark)×\checkmark××Hooker et al.6Image classif.(\checkmark)\checkmark\checkmark××Prior work (PFP)Medical screening(\checkmark)\checkmark\checkmark××This paperMobile image\checkmark\checkmark\checkmark\checkmark\checkmark Summary of representative prior work. \checkmark: addressed; (\checkmark): partially; ×: not addressed. Problem formalization Let be the input space and the binary label space. A classifier parameterized by produces soft outputs and hard predictions . The evaluation distribution over (x, y, a) couples inputs with labels and demographic attribute vector . For each attribute g, denotes the majority group and the minority group. The conditional True Positive Rate is: Normalized objective functions: All component losses are normalized so that weights function as genuine priority coefficients rather than scale-correction factors43: Unlike and , is not bounded above by 1: configurations with or yield , encoding hard-constraint violation directly in the loss magnitude. All DFZ-qualifying configurations have . Constrained multi-objective problem: subject to: Constraint (9) applies per attribute separately, preventing the aggregate from masking per-attribute violations25,44, approximating minimax group fairness26. Threshold rationale is given in the Discussion (Section "Constraint Threshold rationale"). Figure 1 illustrates the trilemma geometry and the DFZ region. Trilemma geometry. Vertices: Accuracy (), Fairness ( per attribute), Efficiency (łe 10 MB, łe 300 ms). Shaded region: Deployment-Feasible Zone (7 configurations). ★: C2 (M7+PFP) Pareto knee point (). Open circles: excluded from DFZ. Scalarization: with . Deployment-Feasible Zone: The Pareto-efficient subset . The knee point: Experimental setup The benchmark follows an eight-stage pipeline that maps each design decision in this paper to a concrete experimental step: (i) dataset construction from the Kaggle Autism Facial Expression collection and locally collected Vietnamese subset (Section "Dataset and partitioning"); (ii) preprocessing through MediaPipe face detection, alignment, and quality filtering (Section “Preprocessing pipeline”); (iii) subject-level 80/10/10 partition performed before any augmentation to prevent identity leakage (Section "Dataset and partitioning"); (iv) backbone selection on a fixed DenseNet-121 to isolate optimization strategy contributions (Section "Backbone selection and shared training protocol"); (v) model training under one of eleven optimization configurations (Section "Eleven optimization configurations"), with weight scheduling governed by the Adaptive Trilemma Weight Scheduler (Section "Adaptive trilemma weight scheduler"); (vi) compression – pruning, quantization, or distillation – applied according to each configuration’s strategy (Section "Eleven optimization configurations"); (vii) multi-environment latency and fairness evaluation across three hardware tiers (Section "Multi-environment evaluation protocol"); and (viii) Pareto-frontier and Deployment-Feasible Zone analysis on the fixed test set (Section "Pareto frontier and sensitivity analysis"). Figures 1, 2, 3, 4, 5, 6 act as visual anchors for the key stages: Fig. 1 defines the trilemma geometry and DFZ; Fig. 2 illustrates the intersectional subgroup distribution; Fig. 3 shows the 3D-aware augmentation pipeline; Fig. 4 details ATWS weight scheduling; Fig. 5 reports per-attribute EOD; and Fig. 6 visualizes the Pareto frontier in normalized objective space. Intersectional subgroup distribution (24 cells: gender × ethnicity × age). Worst-case imbalance 35.47:1 (Male×White vs. Female×Dark n=73, ASD n=19). Red border: two smallest subgroups, which do not reach 90 % per-subgroup sensitivity under any configuration. 3D-aware augmentation pipeline (M7). DECA39 reconstructs FLAME parameters from a 2D input image. (± 5 poses) and (× 2 expressions) are varied; is locked to preserve subject-specific facial structure encoding classification-relevant biomarkers. Each source image yields 10 augmented variants. Applied exclusively to training-fold images (no leakage). ATWS objective weight scheduling across three training phases. Phase I (0–40 % of budget E): accuracy convergence . Phase II (40–70 %): progressive fairness escalation (0.4, 0.4, 0.2). Phase III (70–100 %): joint optimization (0.3, 0.4, 0.3). Override 1: val-EOD . Override 2: val- reset to Phase I for 5 epochs. Equal Opportunity Difference (EOD) across 11 configurations. Dark bars: ; light bars: . Dashed red: hard constraint. ★ C2: lowest EOD in DFZ (, ). Red bars: constraint violation (, DFZ excluded). Pareto frontier in normalized objective space (, , ). Filled circles: DFZ-qualifying configurations. ★: C2 knee point (). Open circle: A2 (Pareto-optimal but size > 10 MB, excluded from DFZ). Gray points: non-Pareto. C2 minimizes Euclidean distance to the ideal point (0, 0, 0). Dataset and partitioning Experiments use a 2,821-image binary classification dataset of children aged 0–6 (1,400 positive class, 1,421 negative) with 24 intersectional subgroups (gender × ethnicity × age group) and worst-case imbalance 35.47:145–48 (Fig. 2). The dataset combines 2,122 images from the Kaggle Autism Facial Expression collection (78.2 % White-presenting) and 699 locally collected images of Vietnamese children. Labels derive from prior clinical diagnosis records49: DSM-5 criteria for the Kaggle subset; M-CHAT-R/F screening and specialist clinical assessment for the Vietnamese subset50. No independent re-labeling was performed by the present authors; label quality is inherited from the source datasets. Label noise, if present, affects all configurations equally and does not bias comparative findings, but may affect absolute performance estimates. Data integrity. Partition: train /val = 281/test = 282 (80/10/10), performed at the subject level before any augmentation to prevent identity leakage. All augmentation – including 3D-aware synthesis via DECA39 – was applied exclusively to training-fold images during cross-validation, never to validation or test sets. DECA-reconstructed variants share the subject-specific shape parameter with the source image; subject-level partitioning ensures no identity information crosses partition boundaries. Preprocessing pipeline Each image was processed through: (i) face detection (MediaPipe Face Mesh, confidence \ge 0.9, pose deviation ); (ii) alignment (inter-ocular axis horizontal: ); (iii) crop to facial bounding box with 5-pixel margin, padded to 224 × 224 with neutral gray (RGB: 128, 128, 128); (iv) quality filter (composite blur-exposure score \ge 0.6). The pipeline retained 2,821 of 2,936 candidate images (96.1 %). Pixels normalized by ImageNet statistics (μ = [0.485, 0.456, 0.406], σ = [0.229, 0.224, 0.225]). Backbone selection and shared training protocol All eleven configurations share DenseNet-12118 pretrained on ImageNet as the fixed backbone, ensuring trilemma outcome differences are attributable to optimization strategy rather than architectural choice11. DenseNet-121 was selected via prior architecture comparison (4 architectures × 3 augmentation methods, 5-fold CV): it achieved the highest average (0.727) and best parameter efficiency (0.0909 /M parameters), outperforming VGG19 (143M), ResNet-50+CBAM51 (25M), MobileNetV231 (3.4M), and EfficientNet-B032 (5.3M). Classification head: global average pooling → Dropout(0.5) → sigmoid. Shared hyperparameters across all configurations: AdamW optimizer; initial LR decaying to via cosine annealing; weight decay 0.01; batch size 32; binary cross-entropy with balanced class weights; early stopping (patience = 10 on val-, max 150 epochs). Hardware: NVIDIA RTX 3080 Ti (12 GB VRAM), CUDA 11.8, PyTorch 1.13.1, random seed 42. Eleven optimization configurations Table 3 presents the complete configuration matrix. Three design principles apply uniformly: (i) the shared backbone ensures comparability; (ii) augmentation-only configurations (A1, A2) use no compression; compression-only (A3, B1–B4) use the A1 data pipeline; combined (C1–C4) apply both; (iii) all hyperparameters are shared to isolate strategy contributions. Table 3Complete configuration matrix.IDGroupStrategyDescriptionNoteA1BaselineBasic 2D-AugRotation , noise, brightness; no compressionLower boundA2Baseline3D-Aware (M7)DECA39: 10 variants/image; fixed; no compressionSize >10 MBA3BaselineStd. PruningMagnitude, 80 % sparsity, no fairness constraintB1CompressPFPPer-neuron EOD check δ =0.08; 80 % compressionB2CompressINT8 Quant.Post-training INT828; 200-img calibrationB3CompressKD→MNv3Distillation30: MNv3-Small4, T=4, λ =0.7B4CompressPFP+INT8Sequential: PFP then INT8; maximum efficiencyC1CombinedM7+Std. Prune3D aug (A2) + standard pruningEOD viol.C2CombinedM7+PFP3D aug (A2) + PFP (B1 settings)Knee pt.C3CombinedM7+INT83D aug (A2) + INT8 quantizationC4CombinedM1+PFPBaseline 2D aug + PFP; data quality ablation Complete configuration matrix. Protected fairness pruning PFP integrates a fairness-aware importance score into the pruning criterion. Algorithm 1 presents the complete procedure. The fairness impact (line 5) uses rather than mean, ensuring no single parameter removal can cause a per-attribute EOD violation masked by averaging – directly enforcing constraint (9). Specifically, mean aggregation would allow a severe fairness violation on one protected attribute (e.g. a sharp increase in ) to be averaged out by minor impacts on another attribute (e.g. ), potentially masking per-attribute violations. The operator prevents this by evaluating the worst-case attribute impact of each parameter removal, as discussed further in the Note below and in the Discussion (Section "PFP Pareto-dominates standard magnitude pruning"). Protected Fairness Pruning (PFP) Hyperparameters: λ = 0.8, B = 10 MB, \varepsilon = 0.10, r = 0.05, k = 30 epochs. PFP is robust to : all values produce DFZ-qualifying models (, Table 5 in Version 1). Table 5Comprehensive fairness metrics. Italics: per-attribute EOD constraint violation (). EOD = Equal Opportunity Difference (measuring TPR parity); DPD = Demographic Parity Difference; SPG = Subgroup Performance Gap (24 cells); MAE = Mean Absolute Error of subgroup across 24 cells. Test set (n = 282).IDConfigSens. (%)Spec. (%) (%) (%) (%) (%) SPG (%)MAE (%)A1Basic Aug0.50648.760.312.116.28.511.326.418.3A23D Aug M70.94694.195.12.13.91.83.27.34.1A3Std. Prune0.87487.488.18.311.66.18.919.813.2B1PFP0.90491.792.32.13.91.83.48.14.6B2INT8 Quant.0.91992.092.62.44.22.03.78.95.0B3KD MNv30.89189.890.53.15.02.64.310.26.1B4PFP+INT80.88789.289.92.34.01.93.58.44.8C1M7+StdPrn0.92192.493.07.910.85.88.417.611.4C2M7+PFP0.93493.594.12.03.71.73.17.84.3C3M7+INT80.93894.094.72.23.81.93.37.54.2C4M1+PFP0.87187.588.22.44.12.03.68.34.7 Comprehensive fairness metrics. Italics: per-attribute EOD constraint violation (). EOD = Equal Opportunity Difference (measuring TPR parity); DPD = Demographic Parity Difference; SPG = Subgroup Performance Gap (24 cells); MAE = Mean Absolute Error of subgroup across 24 cells. Test set (n = 282). Note: For resource-constrained setups, in line 5 of Algorithm 1 can be relaxed to to reduce pruning time by (from h to h), at the expense of strict worst-case per-attribute fairness guarantees. The operator is preferred in the primary configuration because mean aggregation allows a severe fairness violation on one protected attribute (e.g. a sharp increase in ) to be masked by minor impacts on another (e.g. ), violating the per-attribute constraint (9). See also Section "PFP Pareto-dominates standard magnitude pruning" for computational overhead discussion. 3D-aware augmentation pipeline Figure 3 illustrates the M7 pipeline. The M7 pipeline uses DECA39 in the FLAME parametric space . For each minority-group training image: (i) DECA encoder extracts ; (ii) varied ( yaw, 5 poses) and varied (× 2 expressions); (iii) is fixed, preserving classification-relevant structural attributes; (iv) the FLAME renderer produces 10 semantically valid variants with quality filtering. Applied exclusively to training-fold images. Augmentation quality is assessed via downstream task performance and semantic fidelity metrics (FID, LPIPS) rather than pixel-level SSIM/PSNR, consistent with best practices for semantically structured augmentation36,40. Adaptive Trilemma weight scheduler Algorithm 2 and Fig. 4 present ATWS and its weight scheduling. The motivation is the gradient-conflict phenomenon22: introducing fairness regularization before discriminative features stabilize competes with the accuracy gradient23. MGDA-based approaches41 and EPO Search42 address gradient conflict in two-objective settings; ATWS provides a lightweight phased-scheduling alternative for the three-objective case without requiring second-order gradient information. The specific phase boundaries and weights are an original design contribution of this paper. Adaptive Trilemma Weight Scheduler (ATWS) Multi-environment evaluation protocol Inference benchmarking: Latency measured on: E1 – NVIDIA RTX 3080 Ti; E2 – Snapdragon 888/Xiaomi Mi 11; E3 – MediaTek Helio G88/Redmi Note 11 (entry-level SoC, constraint-binding). Mobile inference used TFLite with XNNPACK delegate (float32) or INT8 delegate. Reported values measure model inference only (500 passes after 20-pass warm-up); the full deployment pipeline (face detection ms + alignment ms) adds ms, yielding estimated end-to-end latency of ms for C2 on E3. Fairness metrics: Primary: and (Eqs. 3–4), implementing the Equal Opportunity criterion2. Secondary: Demographic Parity Difference . Supplementary: Subgroup Performance Gap across 24 intersectional cells. We additionally report the Mean Absolute Error across all K=24 intersectional cells, providing a comprehensive average-deviation view alongside the worst-case SPG. Constraint (9) is evaluated per attribute separately25,44. Statistical testing: Training statistics use 5-fold CV (2,258 images; 452 per fold). Test-set 95 % CIs computed by stratified bootstrap resampling (1,000 iterations, n = 282). Pairwise comparisons used the exact permutation test on fold-level values with Holm–Bonferroni correction for 6 pre-specified pairs. Effect sizes reported as Cohen’s d52. Results Main Trilemma benchmark Table 4 presents the trilemma position of all eleven configurations on the fixed test set (n = 282). Seven qualify for the DFZ. Four fail at least one constraint: A1 and A2 exceed the 10 MB size limit; A3 and C1 violate the per-attribute EOD ceiling ( and , respectively), confirming that standard magnitude pruning is fairness-incompatible regardless of upstream augmentation quality. Table 4Main benchmark results: test set (n = 282). with 95 % bootstrap CI. Bold: Pareto knee point (C2). Italics: EOD constraint violation (). *Pareto-optimal but excluded from DFZ (size > 10 MB).IDConfig (95 % CI)Sens. (%) (%) (%)Size (MB)Lat (ms)DFZA1Basic Aug0.506 [0.441, 0.571]48.712.116.227.8218×A23D Aug M70.946 [0.921, 0.971]94.12.13.927.8214×*A3Std. Prune0.874 [0.836, 0.912]87.48.311.66.3187×B1PFP0.904 [0.871, 0.937]91.72.13.96.3187\checkmarkB2INT8 Quant.0.919 [0.888, 0.950]92.02.44.27.1142\checkmarkB3KD MNv30.891 [0.856, 0.926]89.83.15.04.296\checkmarkB4PFP+INT80.887 [0.851, 0.923]89.22.34.03.191\checkmarkC1M7+StdPrn0.921 [0.891, 0.951]92.47.910.86.3187×C2M7+PFP0.934 [0.906, 0.962]93.52.03.76.3187\checkmark(knee)C3M7+INT80.938 [0.911, 0.965]94.02.23.87.1142\checkmarkC4M1+PFP0.871 [0.833, 0.909]87.52.44.16.3187\checkmark Main benchmark results: test set (n = 282). with 95 % bootstrap CI. Bold: Pareto knee point (C2). Italics: EOD constraint violation (). *Pareto-optimal but excluded from DFZ (size > 10 MB). The low baseline performance of A1 () reflects two compounding factors. First, the extreme demographic imbalance (worst-case ratio 35.47:1) combined with class-balanced binary cross-entropy causes conflicting gradient signals: the loss weighting compensates for class imbalance but not for the within-class demographic skew, leading to poor convergence on minority subgroups37. Second, basic 2D augmentation (rotation, noise, brightness) distorts facial morphology patterns relevant to classification without generating semantically valid new samples36. By contrast, A2 () uses 3D-aware DECA augmentation, which generates 10 pose- and expression-varied variants per minority-group image while locking the subject-specific shape parameter that encodes classification-relevant structure. Among DFZ configurations, C2 achieves the highest (0.934) and lowest EOD values (, ), identifying it as the Pareto knee point (). C3 achieves marginally higher (0.938) but slightly higher EOD (); C3 does not dominate C2 given overlapping confidence intervals. B4 (PFP + INT8) achieves the smallest size (3.1 MB) and lowest latency (91 ms), preferred when efficiency is dominant (). Fairness analysis Table 5 and Fig. 5 provide comprehensive fairness metrics. Standard pruning (A3) degrades all fairness metrics relative to A1 (), consistent with Hooker et al.6. C1 (M7 + standard pruning) also fails the per-attribute EOD constraint despite , demonstrating that high-quality upstream augmentation cannot compensate for fairness-unconstrained compression downstream33–35. INT8 quantization (B2) maintains EOD values comparable to PFP ( vs. ), consistent with quantization causing less demographic disparity than magnitude pruning6,33. EOD as defined in Eq. (3) directly implements the Equal Opportunity criterion of Hardt et al.2: it measures the absolute difference in True Positive Rates between majority and minority groups, per protected attribute. The constraint therefore corresponds to requiring that the Equal Opportunity gap does not exceed 10 percentage points for either gender or ethnicity. DPD provides a complementary Demographic Parity view, and SPG captures the worst-case spread across all 24 intersectional cells. Efficiency metrics Table 6 reports efficiency metrics. A key structural finding: A3 and B1 achieve identical efficiency (6.3 MB, 187 ms, ), confirming A3 is strictly Pareto-dominated by B1. The E3 latency spread among DFZ configurations (91–187 ms) provides meaningful differentiation for deployment decision-making. Table 6Efficiency metrics. per Eq. (6), MB, ms. E1: RTX 3080 Ti. E2: Snapdragon 888. E3: Helio G88 (constraint-binding). FLOPs via ptflops.IDConfigSize (MB)FLOPs (G)Lat (ms) Lat (ms) Lat (ms) A1Basic Aug27.82.9117682181.73A23D Aug M727.82.9117682141.71A3Std. Prune6.30.6811431871.26B1PFP6.30.6811431871.26B2INT8 Quant.7.10.728311420.95B3KD MNv34.20.22624960.58B4PFP+INT83.10.17519910.46C1M7+StdPrn6.30.6811431871.26C2M7+PFP6.30.6811431871.26C3M7+INT87.10.728311420.95C4M1+PFP6.30.6811431871.26 Efficiency metrics. per Eq. (6), MB, ms. E1: RTX 3080 Ti. E2: Snapdragon 888. E3: Helio G88 (constraint-binding). FLOPs via ptflops. Pareto frontier and sensitivity analysis Figure 6 visualizes the Pareto frontier in three-dimensional objective space. Five configurations are Pareto-optimal: A2, B1, B2, B3, C2 (and C3 with non-overlapping CI from C2 only on ). A2 anchors the accuracy–fairness vertex () but fails the size constraint. B3 and B4 anchor the efficiency vertex (3.1/4.2 MB, 91/96 ms). C2 and C3 constitute the DFZ-feasible interior. Linear Scalarization (LS) baseline: As a standard MOO reference, we evaluate the classical Linear Scalarization method19,20, which aggregates the three component losses into a single scalar via the convex combination (Eq. 12). Under the canonical uniform-weight setting – which treats accuracy, fairness, and efficiency as equally important – the LS criterion, restricted to DFZ-qualifying configurations, selects C2 as the optimal solution. This LS-optimal choice coincides exactly with the ternary-sensitivity winner (C2 in 24 of 36 weight combinations, ; see below) and with the Pareto knee point defined by the smallest Euclidean distance from to the ideal origin (0, 0, 0), . The convergence of three independent decision criteria – uniform-weight LS, ternary sensitivity, and the Euclidean knee – establishes that C2’s optimality is not an artefact of any single weighting scheme but reflects a structural advantage in the three-dimensional objective space43. Bootstrap stability: C2 was the knee point in 942 of 1,000 bootstrap resamples (94.2 %); C3 in the remaining 5.8 %, reflecting near-identical within overlapping CIs. vs. (Δ = 0.005). In the 5.8 % of resamples where C3 wins, bootstrap fluctuations in the two smallest subgroups (n=19, n=22) occasionally reverse the EOD ordering – a known instability when per-attribute metrics are sensitive to extreme class imbalance. This instability is an acknowledged limitation (see Section “Limitations”), not a finding contested by the main result. Ternary sensitivity: Across 36 weight combinations in (unit-sum), C2 minimized in 24 combinations (66.7 %) – all with and . B3 was optimal in 8 combinations with . C3 in 4 intermediate combinations. The DFZ solution space is characterized by three Pareto-non-dominated options: C2 (balanced), C3 (accuracy-leaning), B3 (efficiency-dominant). Discussion ATWS ablation Table 8 presents pairwise statistical comparisons and Table 9 the ATWS ablation. ATWS consistently improves by 1.3 pp and reduces EOD by 0.6–0.7 pp on both B1 and C2. The improvement is modest (Cohen’s d = 0.51) but consistent across configurations, suggesting ATWS addresses training dynamics rather than model capacity. Phase boundary sensitivity was verified: nine combinations from all satisfy DFZ constraints; no combination degrades by more than 0.4 pp or increases per-attribute EOD by more than 0.3 pp relative to the default (0.4E, 0.7E). Table 7 summarizes key results across these combinations. Table 8Pairwise statistical comparisons. Exact permutation test (n = 5 folds), Holm–Bonferroni corrected. With n = 5, the minimum achievable two-sided p-value is 0.031. : p < 0.10 (marginal, not significant at 0.05).ComparisonCohen’s dSig.InterpretationC2 vs. A3+0.0600.0311.42*Fair compression dominates on all 3 axesC2 vs. B1+0.0300.0630.88†3D aug adds marginal gainC2 vs. C4 (M7/M1)+0.0630.0311.78*3D-aware aug is dominant accuracy driverB1 vs. A3 (equal eff.)+0.0300.0310.82*PFP strictly Pareto-dominates std. pruningC2-ATWS vs. C2-fixed+0.0130.0630.51†ATWS: consistent marginal gainDFZ vs. rest+0.079<0.0012.14**DFZ separates performance tiers Pairwise statistical comparisons. Exact permutation test (n = 5 folds), Holm–Bonferroni corrected. With n = 5, the minimum achievable two-sided p-value is 0.031. : p < 0.10 (marginal, not significant at 0.05). Table 9ATWS ablation. Δ rows: absolute difference (ATWS − fixed). All configurations remain in DFZ after ATWS.IDTrainingSens.(%)(%)(%)gen(%)eth(%)B1Fixed (0.4, 0.4, 0.2)0.89189.62.84.62.34.0B1ATWS0.90491.72.13.91.83.4Δ B1ATWS−fixed+0.013+2.1−0.7−0.7−0.5−0.6C2Fixed (0.4, 0.4, 0.2)0.92192.82.64.32.23.8C2ATWS0.93493.52.03.71.73.1Δ C2ATWS−fixed+0.013+0.7−0.6−0.6−0.5−0.7 ATWS ablation. Δ rows: absolute difference (ATWS − fixed). All configurations remain in DFZ after ATWS. Table 7ATWS phase boundary sensitivity (9 combinations from . All combinations satisfy DFZ constraints. Default marked in bold.Ph.I–IIPh.II–III (C2)(%)(%)DFZ0.3E0.6E0.9292.23.9\checkmark0.3E0.7E0.9312.13.8\checkmark0.3E0.8E0.9302.34.0\checkmark0.4 E0.7 E0.9342.03.7\checkmark0.4E0.6E0.9322.13.8\checkmark0.4E0.8E0.9312.23.9\checkmark0.5E0.6E0.9302.34.0\checkmark0.5E0.7E0.9312.13.9\checkmark0.5E0.8E0.9302.24.0\checkmark ATWS phase boundary sensitivity (9 combinations from . All combinations satisfy DFZ constraints. Default marked in bold. Hyperparameter sharing All eleven configurations used shared hyperparameters to isolate strategy contributions from hyperparameter selection, consistent with controlled-comparison design in systematic MOO benchmarks11. The configuration most affected is B3 (knowledge distillation): T = 4 and λ = 0.7 are defaults from Hinton et al.30 and may understate B3’s performance. Tuned distillation might achieve closer to 0.91–0.93, conservatively biasing the comparison against B3. PFP Pareto-dominates standard magnitude pruning The central empirical finding is that standard magnitude pruning (A3) is strictly Pareto-dominated by PFP (B1) at identical compression ratio. This rests on a causal chain, each link independently supported by prior literature: (E1) Low-magnitude encoding: When a demographic subgroup is underrepresented, gradient signals from that subgroup are proportionally weaker, causing minority-encoding neurons to converge to lower weight magnitudes6,25. (E2) Selective removal under magnitude pruning: Standard pruning removes the lowest-magnitude neurons, non-uniformly targeting minority-group features at a disproportionately higher rate6,35. (E3) PFP breaks the mechanism: PFP’s per-neuron EOD impact check (line 5, Algorithm 1) blocks (E2) by refusing to remove any neuron whose removal would increase per-attribute EOD beyond δ, regardless of its magnitude34. Empirical confirmation: A3 and B1 achieve identical efficiency (6.3 MB, 187 ms, ), yet B1 dominates on (0.904 vs. 0.874, p = 0.031, d = 1.42) and fairness ( vs. ). Furthermore, C1 (M7 + standard pruning, ) fails the EOD constraint while C2 (M7 + PFP) satisfies all constraints with , demonstrating that high-quality augmentation upstream cannot compensate for fairness-unconstrained compression downstream. Computational overhead: PFP incurs approximately 14 hours of pruning time on an RTX 3080 Ti versus 6 hours for standard magnitude pruning (2.3× overhead). This is a one-time offline preparation cost: once the pruned model is obtained, inference latency is completely unaffected and identical to standard pruning (187 ms on E3). For a model deployed over an operational lifetime of months or years, the one-time overhead amortizes to a negligible per-inference cost. A lightweight alternative using mean EOD impact (replacing with in line 5 of Algorithm 1) reduces pruning time to approximately 9 hours ( reduction), with comparable DFZ-qualification rates, at the expense of strict worst-case per-attribute fairness guarantees. This structural trade-off between computational cost and fairness stringency is now explicitly characterized, providing practitioners with flexible deployment options. See the Note at the end of Section "Eleven optimization configurations" and the future work discussion in the Conclusion for details. Constraint threshold rationale The four hard constraints in Eqs. (8)–(11) were selected as follows: : An of 0.85 represents the minimum clinically actionable performance for a screening tool operating alongside human review1,53,54; below this threshold, false-negative rates become unacceptably high for time-sensitive referrals. : The 10 pp EOD threshold corresponds to the Equal Opportunity definition of Hardt et al.2, which is widely adopted as the boundary between acceptable and discriminatory TPR disparity in deployed AI systems27 and binary classification fairness literature3,24. MB: The 10 MB limit is the standard over-the-air (OTA) update size constraint for Android applications under typical cellular data plans in the target deployment region4. ms: The 300 ms interactive latency budget follows established UX guidelines for responsive mobile interfaces5, evaluated on E3 (Helio G88) as the dominant hardware tier in the target deployment region. Sensitivity of DFZ membership to threshold values: Varying by ± 0.02 or by ± 0.02 changes DFZ membership by at most one configuration; the C2 knee-point identity is stable across all tested perturbations. Varying between 8 MB and 12 MB or between 250 ms and 350 ms similarly leaves C2 as the knee point. More aggressive thresholds (e.g. MB) would restrict the DFZ to efficiency-dominant configurations (B3, B4), which is appropriate for IoT deployment contexts11. Sensitivity of sub-weighting. Varying across : DFZ membership is identical in all three cases; C2 remains the knee point; and the PFP-over-standard-pruning dominance holds in all cases since A3 and B1 have identical size and latency ( identical regardless of sub-weighting). Backbone generalization The benchmark fixes DenseNet-121 to isolate strategy contributions from architectural differences11. To assess whether the core findings generalize to lightweight backbones, we additionally evaluated PFP (B1 settings) and the ATWS scheduler on EfficientNet-B032 and MobileNetV3-Small4 under identical hyperparameters. Table 10 reports key trilemma metrics. Table 10Backbone generalization: PFP + ATWS applied to three backbones under B1 settings. EfficientNet-B0 and MobileNetV3-Small results serve as portability evidence; DenseNet-121 is the primary benchmark backbone.BackboneParams(M)(%)Size(MB)DFZDenseNet-121 (primary)7.980.9042.16.3\checkmarkEfficientNet-B05.30.8812.84.9\checkmarkMobileNetV3-Small2.50.8623.43.2\checkmark Backbone generalization: PFP + ATWS applied to three backbones under B1 settings. EfficientNet-B0 and MobileNetV3-Small results serve as portability evidence; DenseNet-121 is the primary benchmark backbone. All three backbones qualify for the DFZ under PFP + ATWS, and in each case PFP-pruned models outperform standard magnitude pruning at the same sparsity level on both and EOD. The PFP Pareto-dominance result (Empirical Observation 1) thus holds across architectural families exhibiting the low-magnitude minority-encoding property. Absolute performance varies with backbone capacity, as expected; the strategy-level conclusions do not depend on DenseNet-121 specifically. Portability of the Trilemma framework The trilemma formalization and DFZ concept are application-agnostic. Three adaptation dimensions govern recalibration to new tasks. (1) Threshold recalibration: The values are derived from the present deployment context. Other contexts require recalibration: lower for triage tools prioritizing recall27; stricter for high-stakes decisions16,17; different for IoT devices or server inference11. (2) Protected attribute adaptation: The benchmark uses two binary attributes. Tasks with additional protected attributes (age, disability, socioeconomic proxies) are accommodated by expanding the attribute vector and redefining computation in Eqs. (3)–(4), with constraint (9) expanded accordingly25,44. (3) Backbone substitution: DenseNet-12118 is fixed to isolate strategy effects. The formalization is backbone-agnostic; Empirical Observation 1 holds for any classifier trained on imbalanced data exhibiting the low-magnitude minority-encoding property6,34. Cross-domain validation on established benchmarks (e.g. CelebA38, FairFace) is a priority for future work. Limitations Small subgroup statistics: The two smallest intersectional cells (Female-Dark, , ; Female-Black, , ) yield approximately 2–3 positive cases in the test split. Per-subgroup statistics for these individual cells are unreliable and are excluded from per-subgroup conclusions. Crucially, EOD is computed at the binary attribute level (gender: Male/Female; ethnicity: White/Asian/Black/Dark), not at the intersectional cell level. The Female-Dark cell contributes to the Female denominator in computation, where Female test samples total approximately 86 images – a sufficient base for the binary comparison. Nonetheless, per-attribute EOD estimates carry bootstrap confidence intervals of ± 1.0–2.5 pp depending on subgroup size, and results for the two smallest cells should be interpreted with caution. Future data collection should prioritize Female×Dark and Female×Black subgroups to enable per-intersectional-cell statistical guarantees25. Dataset scope: Multi-center validation has not been performed. Generalizability to other imaging conditions, camera hardware, age ranges, or geographic populations should not be assumed. Cross-domain validation on established fairness benchmarks (e.g. CelebA38, FairFace) is an explicit priority for future work. Backbone scope: The three-backbone generalization in Table 10 is preliminary: only B1 (PFP) settings are tested, and hyperparameters were not re-tuned per backbone. A full benchmark replication across all 11 configurations with per-backbone hyperparameter optimization is deferred to future work. ATWS mechanism: ATWS is empirically justified through ablation and phase boundary sensitivity analysis. The gradient-conflict hypothesis has not been formally verified for the three-objective case via gradient alignment analysis. The improvement is consistent but modest (1.3 pp , Cohen’s d = 0.51, p = 0.063). Compression coverage: The benchmark evaluates four compression families. Structured pruning, lottery ticket methods, mixed-precision quantization29, and NAS-based compression5,9 are not included. Including these would expand the Pareto frontier but is unlikely to change the central finding. Thermal effects on latency: All E3 latency measurements were under single-inference conditions. The 300 ms constraint provides a 60 % margin over C2’s 187 ms; sustained-inference profiling under thermal throttling should be validated before production deployment. Robustness to input distribution shifts: Robustness testing under input distribution shifts (e.g. Gaussian noise injection, brightness or contrast perturbations) and out-of-distribution inputs has not been conducted in the current study. While our optimal configuration maintains a 60 % latency safety margin against thermal throttling (C2: 187 ms vs. ms), extensive robustness validation remains a critical future step before clinical deployment. Conclusion We formalized the Accuracy–Fairness–Efficiency Trilemma in mobile image classification as a constrained multi-objective optimization problem19,20 with normalized component losses and hard deployment constraints, making three distinctions explicit: objectives vs. constraints, aggregate vs. per-attribute fairness2,25, and Pareto optimality vs. deployment admissibility. The resulting Deployment-Feasible Zone provides a principled, application-agnostic method to identify configurations that are simultaneously Pareto-optimal and deployment-feasible. A benchmark of eleven optimization configurations yields three actionable findings. (1) C2 (3D-aware augmentation + PFP) is the Pareto knee point (, , 6.3 MB, 187 ms), confirmed in 94.2 % of bootstrap resamples and optimal in 66.7 % of ternary weight scenarios. (2) Standard magnitude pruning is strictly Pareto-dominated by PFP at identical compression ratio — a result supported by Empirical Observation 1 and the fairness-under-compression literature6,34,35, with statistically significant empirical confirmation (p = 0.031, Cohen’s d = 1.42). (3) ATWS provides consistent marginal improvements (+1.3 pp , -0.6–0.7 pp EOD) over fixed-weight training22,23, compatible with any fairness-constrained optimization strategy. The benchmark design — fixed backbone18, fixed partition, leak-free augmentation39, multi-environment latency, bootstrap CI, and ternary sensitivity analysis20,21 — establishes a reproducible methodology extensible to other mobile classification tasks. Code, pretrained weights for all eleven configurations, and the fixed test-set partition will be released upon acceptance. Cross-domain validation on established benchmarks (e.g. CelebA, FairFace) with different demographic structures and classification tasks would strengthen the portability claim and is a direct avenue for future work. Future research will explore adaptive phase search mechanisms, such as gradient alignment analysis, to dynamically optimize the ATWS boundaries during training. Additionally, a formal, large-scale comparison of versus aggregation strategies in PFP under varying dataset imbalance ratios is a valuable direction for future investigation, as this comparison could reveal the conditions under which the mean-based lightweight variant achieves comparable fairness guarantees to the max-based formulation. Acknowledgements This work was supported by Lac Hong University, Vietnam under grant LHU-RF-TE-26-03-21. Author contributions VT Tran, TH Lam, and NT Do: Writing - review & editing, Writing - original draft, Visualization, Supervision, Project administration, Methodology, Investigation, Data curation, Conceptualization. VT Tran and TT Huynh: Review & editing, Writing, Visualization, Supervision, Project administration, Methodology, Investigation, Data curation, Conceptualization. All authors read and approved the final manuscript. Data availability Implementation code for all eleven configurations, pretrained model weights, and the fixed test-set partition are publicly available at https://github.com/dayhocplc/data-images. The Vietnamese children dataset is available upon reasonable request, subject to institutional data sharing agreement. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement Implementation code for all eleven configurations, pretrained model weights, and the fixed test-set partition are publicly available at https://github.com/dayhocplc/data-images. The Vietnamese children dataset is available upon reasonable request, subject to institutional data sharing agreement.