ML Security PapersAlém do Desempenho: Um Estudo da Confiabilidade de Detectores de Deepfakes
Lucas Lopes 1, Rayson Laroca 2,1, André Grégio 1
Published on arXiv
2601.08674
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
Analysis of five SOTA deepfake detectors reveals significant progress but critical limitations across transferability, robustness, interpretability, and computational efficiency dimensions.
Deepfakes are synthetic media generated by artificial intelligence, with positive applications in education and creativity, but also serious negative impacts such as fraud, misinformation, and privacy violations. Although detection techniques have advanced, comprehensive evaluation methods that go beyond classification performance remain lacking. This paper proposes a reliability assessment framework based on four pillars: transferability, robustness, interpretability, and computational efficiency. An analysis of five state-of-the-art methods revealed significant progress as well as critical limitations.
Key Contributions
- A reliability assessment framework for deepfake detectors structured around four pillars: transferability, robustness, interpretability, and computational efficiency
- A consolidated global reliability score enabling standardized cross-method comparison beyond traditional classification metrics
- Analysis of five state-of-the-art deepfake detectors using the proposed framework, exposing critical gaps and trade-offs
🛡️ Threat Analysis
The paper directly addresses deepfake detection — AI-generated image/video content detection — which is squarely ML09 (Output Integrity). The framework evaluates five SOTA deepfake detectors on dimensions beyond accuracy, contributing evaluation methodology for AI-generated content authentication systems.