Explainability-Guided Deepfake Detection for High-Fidelity Facial Edits

Abstract

Recent advances in vision-language models (VLMs) have enabled highly realistic, identity-preserving facial edits that lack visible artifacts exploited by conventional deepfake detectors. This raises a critical question for multimedia forensics: do models capable of generating such content also possess reliable forensic reasoning? We first show that state-of-the-art VLMs, despite strong performance on standard benchmarks, fail to consistently detect high-fidelity semantic edits, exhibiting viewpoint sensitivity, prompt-dependent behavior, and reliance on superficial cues rather than manipulation-aware visual evidence. To rigorously study this setting, we introduce Side-VLM, a multi-view deepfake dataset comprising minimal facial edits with pixel-level manipulation masks, enabling joint evaluation of detection accuracy and explanation fidelity. Our analysis further reveals that standard deep learning detectors, although achieving high accuracy after fine-tuning, often succeed for the wrong reasons, attending to background or incidental regions rather than manipulated facial attributes. Motivated by these findings, we propose an explainability-guided training framework that enforces right prediction for the right reason by explicitly aligning class activation maps with ground-truth manipulation masks during learning. Extensive experiments demonstrate that the proposed approach significantly improves detection robustness and explanation faithfulness under challenging viewpoints and common image perturbations. Together, our results highlight the limitations of language-driven forensic reasoning and establish explanation-aligned visual learning as a necessary step toward trustworthy deepfake detection in the era of high-fidelity generative models. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2027.