OrEdge: Efficient Multi-Modal Anomaly Detection in Distributed Software Systems via Orthogonal-Domain Learning

Amr M. Zaki, Farhoud Jafari Kaleibar, Honggeun Ji, Komal Sarda, Marin Litoiu · 2026

Read OrEdge: Efficient Multi-Modal Anomaly Detection in Distributed Software Systems via Orthogonal-Domain Learning by Amr M. Zaki, Farhoud Jafari Kaleibar, Honggeun Ji, Komal Sarda, Marin Litoiu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2608-00309.txt.

Introduction

Zaki and co-authors introduce OrEdge for anomaly detection in distributed software systems. Existing tools lean on attention- and graph-based designs. Those designs are computationally expensive and large. OrEdge uses orthogonal-domain temporal representations instead. It jointly reads logs, metrics, and traces from a system.

Its OrEdgeCore module rebuilds signals in an orthogonal domain. That captures recurring temporal patterns while damping transient noise. Tests ran on three real-world microservice data sets. Detection stayed competitive with far larger rival models. The reconstruction model holds at most 9.6K parameters, versus 20K to 143K for rivals. On Raspberry Pi hardware, inference ran sub-second. Latency fell by over an order of magnitude versus prior tools. Ablations back each design component.

Catalog record

CREATORAmr M. Zaki, Farhoud Jafari Kaleibar, Honggeun Ji, Komal Sarda, Marin Litoiu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2608.00309

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

OrEdge: Efficient Multi-Modal Anomaly Detection in Distributed Software Systems via Orthogonal-Domain Learning Amr M. Zaki, Farhoud Jafari Kaleibar, Honggeun Ji, Komal Sarda, Marin Litoiu Abstract We introduce Orthogonal-Edge (OrEdge), a lightweight framework for real-time anomaly detection in multi-modal distributed software systems. Unlike existing approaches that rely on computationally expensive attention- and graph-based architectures, OrEdge leverages orthogonal-domain temporal representations to achieve accurate anomaly detection with substantially lower computational complexity and model size. It jointly analyzes heterogeneous monitoring data, including logs, metrics, and traces, to identify abnormal software behavior, capture temporal dependencies, and reduce redundancy across observability signals. At its core, OrEdge incorporates OrEdgeCore, a lightweight orthogonal-domain reconstruction module that captures recurring temporal patterns while suppressing transient variations. Evaluated on three real-world microservice datasets (MSDS, SN, and TT), OrEdge achieves competitive detection performance while reducing the reconstruction model size to at most 9.6K parameters, compared with 20K–143K parameters in existing methods. This compact design enables efficient deployment on resource-constrained edge devices: on Raspberry Pi platforms, OrEdge achieves sub-second inference and reduces inference latency by over an order of magnitude compared with existing approaches. Extensive ablation studies, sensitivity analyses, orthogonal basis evaluations, and qualitative case studies further validate the effectiveness of each design component. Overall, OrEdge demonstrates that orthogonal-domain temporal modeling provides an effective alternative to computationally intensive attention- and graph-based architectures, achieving a favorable balance between detection accuracy and computational efficiency for real-time multi-modal anomaly detection in edge environments. The code is available at https://github.com/theamrzaki/MicroService_Twin_Original . I Introduction The widespread adoption of distributed computing and microservices has enabled scalable, flexible, and resilient software systems by decomposing applications into loosely coupled, independently deployable services [ 1 , 2 , 3 ] . However, this architectural shift also increases operational complexity, making timely anomaly detection essential for maintaining system reliability and availability [ 4 ] . Anomalies caused by software defects, hardware failures, network disruptions, or security incidents can rapidly propagate across service dependencies if left undetected [ 5 ] . As edge computing continues to expand across IoT devices [ 6 ] , connected vehicles [ 7 , 8 , 9 ] , UAVs [ 10 ] , and mobile systems [ 11 ] , anomaly detection algorithms must provide high detection accuracy while operating under strict computational and memory constraints. Recent multi-modal anomaly detection approaches [ 1 , 12 , 2 , 3 , 13 , 14 , 15 ] improve detection accuracy by jointly modeling logs, metrics, and traces using graph neural networks, Transformers, and large language models. However, these architectures introduce substantial computational and memory overhead, limiting their deployment in real-time and resource-constrained environments. Orthogonal-domain methods [ 16 , 17 , 18 , 19 , 20 ] provide a promising direction for efficient temporal modeling by learning compact representations with reduced redundancy. Nevertheless, existing approaches mainly target generic or single-source time-series and remain largely unexplored for multi-modal software observability. Conversely, lightweight anomaly detectors [ 21 , 22 ] enable efficient edge deployment but are limited to individual telemetry sources, preventing them from capturing cross-modal dependencies among heterogeneous monitoring signals. Therefore, developing an accurate multi-modal anomaly detection framework that maintains low computational cost for edge deployment remains an open challenge. To address these limitations, we propose Orthogonal-Edge (OrEdge) , a lightweight framework for multi-modal anomaly detection in distributed software systems. OrEdge jointly analyzes logs, metrics, and traces by projecting heterogeneous monitoring signals into orthogonal temporal representations that reduce redundancy while preserving informative temporal patterns. At its core, OrEdge introduces OrEdgeCore , a lightweight orthogonal reconstruction module that models recurring temporal behavior using linear projections and linear attention instead of computationally expensive Transformer or graph-attention architectures. This design enables efficient multi-modal anomaly detection while maintaining low computational and memory overhead. We evaluate OrEdge on three real-world microservice datasets (SN, TT, and MSDS) against representative anomaly detection baselines. OrEdge achieves competitive detection performance while requiring only up to 9.6K reconstruction parameters, substantially reducing model complexity compared with existing graph-based and multi-modal anomaly detection approaches. Furthermore, OrEdge enables practical deployment on resource-constrained edge platforms, including Raspberry Pi 3 and Raspberry Pi 5, achieving sub-second inference and reducing latency by over an order of magnitude compared with existing methods. These results demonstrate that orthogonal-domain temporal modeling provides an effective alternative to computationally intensive graph- and attention-based architectures for real-time multi-modal anomaly detection. This work is guided by the following research questions: • RQ1: How does OrEdge compare with state-of-the-art multi-modal anomaly detection approaches in terms of accuracy? • RQ2: How does OrEdge enable efficient edge deployment in terms of model complexity and inference latency? • RQ3: What are the contributions of OrEdge components and orthogonal projection spaces to anomaly detection performance? • RQ4: How sensitive is OrEdge to key design parameters across different multi-modal workloads? • RQ5: A qualitative analysis of OrEdge’s anomaly detection behavior compared with existing methods. The main contributions of this work are summarized as follows: • We propose OrEdge , a lightweight orthogonal-domain framework for multi-modal anomaly detection that jointly models logs, metrics, and traces through compact temporal representations. • We introduce OrEdgeCore , an orthogonal reconstruction module that substantially reduces model complexity while preserving competitive anomaly detection performance. • We conduct comprehensive experiments on three real-world microservice datasets, evaluating detection accuracy, computational efficiency, orthogonal basis selection, architectural ablations, parameter sensitivity, and deployment performance. • We demonstrate practical edge deployment on Raspberry Pi platforms, showing substantial reductions in latency. TABLE I: Comparison of Anomaly Detection Approaches for Multi-Modal Data Method/Year Telemetry Dependency Modeling Architecture Edge Evaluation Domain Classical and Lightweight Methods Single-modality statistical or neural approaches [ 23 ] (DeepLog 2017) Logs ✗ LSTM – Time [ 24 ] (AlertRank 2022) L+M+A ✗ XGBoost – Time [ 21 ] (FADSD 2025) Single TS ✗ Spectrum Score RPi 4B (2GB), STM32 ( < 1MB SRAM) Freq. [ 22 ] (LFTSAD 2025) Single TS ✗ Dual-branch MLP RPi 4B (2GB), Jetson NX (8GB) Time Multi-Modal Methods Graph-based Fusion Methods [ 1 ] (AnoFusion 2023) L+M+T ✓(GAT Graph) GAT – Time [ 12 ] (Eadro 2023) L+M+T ✓(Service Graph) GAT – Time [ 14 ] (MUAD 2026) L+M+T ✓(Dynamic Graph) GAT + Confidence Net (Supervised) – Time [ 25 ] (MGFusion 2026) L+M+T ✓(Grassmann Graph) Graph Attention – Grassmann Transformer and Cross-Modal Fusion Methods [ 2 ] (MSTGAD 2023) L+M+T ✓(Trace Graph) Spatial-Temporal Transformer – Time [ 3 ] (LLM4MST 2024) L+M+T ✓(Trace Graph) GPT-2 – Time [ 13 ] (Art 2024) L+M+T ✗ Transformer + GRU – Time [ 26 ] (Medicine 2024) L+M+T ✗ BERT + Attention – Time [ 15 ] (MAD-CMC 2025) L+M ✓(FC Graph) Cross-modal LSTM – Time Frequency / Orthogonal Domain Methods Compact forecasting and reconstruction models [ 18 ] (FourierGNN 2023) Single TS ✓(FC Graph) Linear Layers – Freq. [ 16 , 17 ] (FITS/FilterNet 2024) Single TS ✗ Linear Layers – Freq. [ 20 ] (FreDF 2025) Single TS ✗ Linear Layers – Freq.+Time [ 19 ] (OLinear 2025) Single TS ✗ NormLinear – Orthogonal [ 27 ] (FFAD 2025) L+M ✓(Fourier Graph) Linear Layers – Freq. OrEdge (Ours) L+M+T ✗(Flattened Graph) Linear Projection + Linear Attention + NormLinear RPi 3 (1GB), RPi 5 (16GB) Orthogonal L: Logs, M: Metrics, T: Traces, A: Alerts, TS: Time

Read the complete text