Researchers have investigated the transferability of adversarial examples across different deepfake detectors, finding that compatibility between source and target models significantly impacts attack success. Their controlled evaluation across 60 detectors revealed that shared backbones, architecture families, pretraining regimes, or training data lead to higher transfer rates. The study also highlighted that averaging attack success rates across multiple sources can underestimate a target model's vulnerability, with a multi-source oracle achieving a much higher success rate when specific compatibility factors were excluded. AI
IMPACT Findings establish source--target compatibility and source-model selection as critical for credible transfer-based black-box robustness evaluation.
RANK_REASON Academic paper detailing a controlled evaluation of adversarial transferability in deepfake detection. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
