Key Explainable AI Techniques Used in Deepfake Detection Explainability in deepfake detection is achieved through a combination of model-level and post-processing techniques. These methods are designed to highlight which inputs or features influenced a model’s decision and how strongly they contributed. Feature Attribution Methods Feature attribution identifies which input parts most influenced the model’s decision. It connects predictions to specific signals rather than treating the output as a black box. Highlights key regions in video or important frequency ranges in audio Shows which features contributed most to the prediction Supports validation of whether the model focused on meaningful signals In practice, it can point to subtle facial inconsistencies, lip-sync mismatches, or unusual speech patterns. This makes it easier to review why a piece of media was flagged during analysis or audits. Saliency Maps and Visual Heatmaps Saliency maps visually highlight areas in an image or video frame that influenced the model’s decision. They make detection results easier to interpret, especially for non-technical users. Highlights important regions in visual inputs Identifies facial artifacts or blending issues Provides visual support for detection results For example, they may emphasize eyes, mouth, or edge regions where synthetic edits are often detected. These signals are typically used alongside other methods since they show influence, not absolute proof. Attention Mechanisms in Multimodal Models Attention mechanisms assign importance to different inputs in models that process audio and video together. They show which modality influenced the decision most. Multimodal layering combines explanations from audio, video, and other inputs into one view. It helps show how different signals interact in a single decision. Weighs audio, video, or text inputs based on relevance Shows which signal impacted the prediction most Can identify mismatches across modalities Merges explanations from multiple data sources Identifies cross-modal inconsistencies Provides a unified view of model reasoning For example, video may appear natural while audio shows synthetic traits. Combining both provides insight into where the manipulation is detected across modalities. For instance, a system may rely more on audio cues for speech anomalies and more on video cues for facial inconsistencies. This helps explain how different signals contribute to one decision. Model-Agnostic Explanation Techniques (LIME and SHAP) These methods explain predictions without depending on a specific model type. They analyze how changes in input affect output results. Measures how individual features influence predictions Works across different model architectures Helps compare feature importance across systems They may show how changes in pitch, frame consistency, or facial motion affect detection results. This makes them useful during model testing and evaluation. 

Source: https://www.resemble.ai/resources/explainable-ai-in-deepfake-detection-techniques