FlorentTariolle/Opportunistic-Adversarial-Attacks
Hybrid query-optimization framework for black-box adversarial attacks that dynamically switches from untargeted exploration to targeted exploitation
What's novel
Hybrid query-optimization framework for black-box adversarial attacks that dynamically switches from untargeted exploration to targeted exploitation
Code Analysis
12 files read · 4 roundsImplements Opportunistic Target Selection (OTS), a technique that modifies untargeted black-box adversarial attacks (SimBA, Square, Bandits) by monitoring which non-true class is leading and switching to a targeted objective when that class stabilizes, improving query efficiency.
Strengths
The OTS mechanism is a genuinely novel and well-motivated contribution to adversarial attacks, with thorough benchmarking across multiple models, attacks, and modes. The Square implementation cleverly subclasses torchattacks.Square to inject OTS hooks without reimplementing the core algorithm.
Weaknesses
The OTS stability-tracking logic is duplicated across all three attack implementations rather than being extracted into a shared mixin or strategy pattern. There are no unit tests at all — the project relies entirely on benchmark scripts for validation.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
180
Contributors
1
Files
715
Active weeks
8
Repository
Language
Python
Stars
1
Forks
0
License
MIT