HyFuzz: A Hybrid AI-Enhanced Vulnerability Detection Framework
As Large Language Models (LLMs) fundamentally reshape the landscape of automated software testing, the security community faces a critical architectural dilemma: how to reconcile the massive execution throughput of traditional fuzzing with the high latency reasoning of neural networks. While generative agents promise to penetrate complex logic states inaccessible to random mutation, their integration imposes a prohibitive computational tax that remains insufficiently quantified in existing literature. To empirically resolve this tension, this thesis presents HyFuzz, a rigorous ablation framework designed to deconstruct the „efficiency intelligence paradox” under strictly controlled resource constraints.Moving beyond simple performance benchmarking, the methodology implemented a controlled „level playing field“ evaluation across five distinct generative strategies. By targeting the structural dichotomy between the rigid, stateless constraints of HTTP and the deep, stateful logic of FTP, the study isolates the specific marginal utility of AI-driven initialization and mutation. Quantitative results demonstrate that while neural integration incurred a 95.6 throughput penalty, it achieved a 37.5 increase in unique fault triggering diversity in the FTP protocol compared to the baseline. These findings delineate the operational boundaries of neural testing, advocating for a hybrid paradigm where generative reasoning is not applied monolithically, but selectively deployed as a precision instrument for complex state navigation.
Zeit: 10:15 Uhr
Ort: Raum 04.137, Martensstr. 3, Erlangen
oder
Zoom-Meeting beitreten:
https://fau.zoom-x.de/j/68350702053?pwd=UkF3aXY0QUdjeSsyR0tyRWtLQ0hYUT09
Meeting-ID: 683 5070 2053
Kenncode: 647333