Overview
Field: Computer Vision (CV) Authors: Christos Korgialas, Gabriel Lee Jun Rong, Dion Jia Xu Ho Published: 2025-07-16 arXiv: 2507.12500
Summary
The reliability of deepfake detectors frequently degrades under black-box adversarial transfer, as these models often rely on fragile, architecture-dependent forensic cues. Existing transfer attacks often lack semantic awareness and struggle to maintain effectiveness under strict no-query constraints, particularly when perturbations are transferred from convolutional surrogates to transformer-based targets.
To address these limitations, the paper introduces ARMOR++, a robust multi-agent framework designed for high-transferability deepfake evasion:
- Qwen2.5-VL (VLM) supplies spatial semantic priors.
- Qwen3 (LLM) orchestrates primitive selection, adaptive hyperparameter reparameterization, and entropy-regularized perturbation mixing.
- Five complementary primitives are integrated, spanning dense optimization, saliency methods, spatial transformations, frequency-domain perturbations, and patch-structure modifications, effectively targeting heterogeneous inductive biases.
- Rigorous evaluation on the AADD-2025 benchmark shows ARMOR++ significantly outperforms existing agentic and non-agentic baselines in both low-quality and high-quality image scenarios.
- Statistical analysis confirms substantial improvements in blind-target attack success rate (ASR) over state-of-the-art agentic baselines.
- ARMOR++ shows further advantages against non-agentic baselines and robust defense configurations.
- The results reveal significant residual reliability gaps in current deepfake detector deployments and demonstrate the effectiveness of agentic orchestration in uncovering latent vulnerabilities.