Fuzz Smarter, Not Harder: Towards Greener Fuzzing with GreenAFL

Ayse Irmak Ercevik, Aidan Dakhama, Melane Navaratnarajah, Yazhuo Cao, Leo Fernandes · 2025

Read Fuzz Smarter, Not Harder: Towards Greener Fuzzing with GreenAFL by Ayse Irmak Ercevik, Aidan Dakhama, Melane Navaratnarajah, Yazhuo Cao, Leo Fernandes (2025) 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-2510-25665.txt.

Introduction

Ayse Irmak Ercevik led four co-authors on greener fuzzing. Aidan Dakhama, Melane Navaratnarajah, and Yazhuo Cao helped. Leo Fernandes closes the five-author team. Their 2025 paper presents GreenAFL, an energy-aware fuzzing approach. Fuzzing is a key search-based software testing technique.

Current grey-box fuzzers focus primarily on coverage maximization. AFL++ exemplifies that coverage-first method to fuzzing. Energy costs of exploring execution paths are ignored. Energy-aware corpus minimization is the first modification. It considers power when reducing initial corpora. Energy-guided heuristics direct mutation toward quick inputs. High-coverage, low-energy inputs are preferred by heuristics. An ablation study compares vanilla AFL++ against variants. Each modification part contribution is tested separately. Highest coverage and lowest energy came with modifications.

Catalog record

CREATORAyse Irmak Ercevik, Aidan Dakhama, Melane Navaratnarajah, Yazhuo Cao, Leo Fernandes
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE15 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-25665

Browse the full library · About the Tower of Babel

Full text

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

Fuzz Smarter, Not Harder: Towards Greener Fuzzing with GreenAFL Ayse Irmak Ercevik, Aidan Dakhama, Melane Navaratnarajah, Yazhuo Cao, Leo Fernandes Abstract Fuzzing has become a key search-based technique for software testing, but continuous fuzzing campaigns consume substantial computational resources and generate significant carbon footprints. Existing grey-box fuzzing approaches like AFL++ focus primarily on coverage maximisation, without considering the energy costs of exploring different execution paths. This paper presents GreenAFL , an energy-aware framework that incorporates power consumption into the fuzzing heuristics to reduce the environmental impact of automated testing whilst maintaining coverage. GreenAFL introduces two key modifications to traditional fuzzing workflows: energy-aware corpus minimisation considering power consumption when reducing initial corpora, and energy-guided heuristics that direct mutation towards high-coverage, low-energy inputs. We conduct an ablation study comparing vanilla AFL++, energy-based corpus minimisation, and energy-based heuristics to evaluate the individual contributions of each component. Our evaluation shows up to 7.4% lower energy usage and 7.1% lower throughput while maintaining or improving coverage, with best-case coverage gains of 2.6%. 1 Introduction Fuzzing has emerged as a powerful greedy search-based technique for discovering bugs in complex software systems, serving as a common search-based technique in industrial testing pipelines; however, it comes with significant computational and energy costs. Systems such as OSS-Fuzz [ 11 ] run continuous fuzzing campaigns using substantial resources, consequently producing large carbon footprints. Similarly, fuzzing is increasingly applied to large-scale and energy-intensive systems such as operating systems [ 3 , 13 ] , quantum system simulators [ 2 ] , and system simulators [ 5 , 6 , 8 ] . As automated testing scales, it becomes increasingly important to manage the energy usage of such systems. Grey-box fuzzing techniques such as AFL++ use coverage feedback to guide input mutation towards more effective test cases. Whilst this coverage-guided approach improves the efficiency of discovering new execution paths, it fails to consider the energy consumption associated with exploring these paths. Consequently, running fuzzing campaigns remains an energy-intensive process. Much of the existing work on reducing the emissions of fuzzing focuses on stopping criteria [ 9 ] , or in the case of GreenBench , improving the emissions of the benchmarks used to evaluate fuzzers [ 12 ] . Yue et al.’s EcoFuzz reduces redundant test case generation through optimised scheduling with an Adversarial Multi-Armed Bandit model, achieving higher coverage from fewer executions [ 14 ] . Similarly, Lyu et al.’s SLIME introduces program-sensitive energy allocation to adaptively distribute fuzzing effort across seeds rather than measuring actual power use [ 10 ] . Unlike our approach, neither EcoFuzz nor SLIME measure actual energy usage, but instead optimise seed scheduling to indirectly reduce waste. We propose GreenAFL , a modification to afl-cmin that considers the system energy usage of each input when minimising the initial corpus. Further, we extend the heuristics used by AFL++ to guide fuzzing towards inputs that achieve high coverage whilst maintaining low CPU and memory energy consumption. To the best of our knowledge, this is the first work that directly incorporates energy usage into fuzzing and minimisation heuristics. In our evaluation, energy-aware corpus minimisation consistently delivered the best balance of reduced energy (up to 7.4%) and maintained or improved coverage (up to 2.6%), while energy-guided fuzzing heuristics showed promising benefits but require further refinement. Together, these results demonstrate that direct integration of energy awareness into fuzzing workflows is both feasible and impactful. Beyond sustainability, reducing energy translates directly into lower infrastructure costs, making greener fuzzing attractive for industrial-scale deployments. Availability. GreenAFL , and documentation demonstrating its generalisability, reproducibility materials, and experimental results, are available at [ 1 ] . 2 GreenAFL GreenAFL extends the AFL++ fuzzing framework by integrating energy awareness into corpora minimisation ( green-cmin ) and energy-aware fuzzing heuristics ( green-afl ). Our approach leverages Intel’s hardware-based power monitoring via cppjoules [ 4 ] . Our implementation is generic and modular, utilising the LD_PRELOAD library that wraps the system under test to record energy usage to ensure portability. While our implementation is Intel-specific, the preload-based mechanism can be adapted to other architectures (e.g. AMD or ARM CPUs) and extended to NVIDIA GPUs, where cppjoules already provides support [ 4 ] . 2.1 Energy-Aware Corpus Minimisation AFL’s cmin reduces the initial corpus by discarding redundant inputs, keeping only those that contribute new coverage. GreenAFL augments this process by also considering the energy consumption of each input. Specifically, we retain inputs that maximise coverage while minimising power cost – balancing effectiveness with efficiency; this can reduce the emissions required for a test, as well as serve to reduce running costs, especially in the context of large scale long running campaigns. This ensures that the subsequent fuzzing loop operates on inputs that are both coverage-rich and energy-efficient. GreenAFL runs afl-showmap once per seed to record which program edges the seed reaches, while measuring the total energy for that run. For each edge, we keep the seed that has the lowest energy cost. It is important to note that this measurement reflects the energy consumed during the execution of each seed by the fuzzer, rather than the long-term cost of the resulting test archive. 2.2 Energy-Guided Fuzzing Loop In traditional AFL++ , the fuzzing loop schedules mutant generation according to new coverage. GreenAFL extends this loop with energy-aware heuristics that modify the allocation of mutation cycles. These heuristics aim to bias the search towards executions with higher “coverage-per-watt”. Figure 1 : Overview of GreenAFL ’s energy-guided fuzzing loop. The green boxes with bold text highlight where our energy-aware heuristics are applied, ( B , energy-aware score computation) and ( D , airtime scheduling). Airtime Scheduling via Energy-Scaled Performance Score In the AFL++ fuzzing loop ( fig. 1 , D ) each seed has a perf_score computed which determines how much fuzzing effort is allocated and represents a combination of factors – such as execution speed and input size – which is used to determine the number of mutation cycles. We modify perf_score to be energy - aware by applying a scale factor inversely proportional to the energy used during the execution of the seed. The multipliers are derived from each seed’s CPU and memory energy relative to the campaign’s global minimum and maximum energy cost, which we map to the range of 5\times to \frac{1}{5}\times . This prioritises low energy seeds, allowing for more mutation cycles ( fig. 1 , E ). Energy-Aware Favoured Seed Selection AFL++ assigns a single champion seed to each edge transition in the coverage bitmap ( fig. 1 , B ). Whenever a seed is executed and hits an instrumented edge, its score is compared against the current champion’s score. If the new seed has a lower score, it replaces the current champion for that edge. We make this selection energy-aware by applying an energy-based scale factor in the range of \frac{4}{5}\times to \frac{5}{4}\times . The multiplier is obtained by a logarithmic normalisation of the seed’s total energy cost relative to the campaign’s minimum and maximum energy costs. This prioritises low-energy seeds and penalises high-energy seeds, marking low-energy seeds as favoured ( fig. 1 stage C ). Favoured seeds are then prioritised in the fuzzing schedule and selected more often for mutation. 2.3 Experimental Setup We

Read the complete text