InfoResearchPeer-reviewed
Toward Interpreting and Boosting Adversarial Transferability via Transferability Error Minimization
- Published
- Record updated
Summary
The paper introduces Transferability Error, the gap in prediction-change risk between an adversarial example and the optimal example in the adversarial region, and links it to the surrogate model's maximum prediction discrepancy. Based on this bound, the authors propose Transferability Error Minimization (TEM), which adds a differentiable regularizer to the standard loss. On ImageNet-compatible and CIFAR-10, TEM reportedly outperforms baselines by up to 5% in average attack success rate against normally trained, adversarially trained and defended models.
Related items
- MediumGHSA-25rm-9wvm-m38v: falcon-rust: Discrete-Gaussian sampler runs at 2^-33 precision, far below Falcon's requirement (reference KAT vectors no longer pass)Similar attack · GitHub Advisory Database
- InfoRECLAIM: Robot-Vision-Informed Adversarial Patch Defense via Embodied Attention FeedbackSimilar attack · Arxiv (cs.RO + cs.CV security)
- LowExecGap: Exploiting Command-Execution Gap for Patch Attacks on Vision-Language-Action ModelsSimilar attack · Arxiv (cs.RO + cs.CV security)
- InfoConfGuard: a physics-grounded reference monitor for predictive-maintenance large-language-model agentsSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
- MediumGemini AI agents repeatedly bypassed safeguards to attack real websiteSimilar attack · BleepingComputer