Why Fake ? Unveiling the Semantic Vocabulary of Deepfake Detectors

Technical Summary: Unveiling the Semantic Vocabulary of Deepfake Detectors

Problem Statement
Deepfake (DF) technology poses a critical threat to information integrity, necessitating robust detection methods. However, the prevailing approach in deepfake detection relies on binary classification (real vs. fake), which lacks the transparency and justification required for high-stakes applications such as legal proceedings. While Explainable Deepfake Detection (XDFD) has emerged to address this, existing techniques suffer from significant limitations: spatio-temporal localization methods often produce coarse visual evidence (e.g., highlighting the entire face rather than specific artifacts), and textual explanation methods frequently lack accurate spatial grounding or fail to reflect the underlying evidence used by the detector. Furthermore, many current XDFD approaches require architectural modifications, auxiliary training objectives, or human annotations, creating trade-offs between interpretability and predictive performance.

Methodology
This work proposes a post-hoc Explainable AI (XAI) framework that analyzes state-of-the-art black-box deepfake detectors without modifying their architecture or requiring retraining. The core technical approach utilizes Encoding-Decoding Direction Pairs (EDDP), a technique grounded in the linear representation hypothesis.

1. Concept Extraction: The method assumes that concepts (features indicative of real or fake images) are encoded within the model's latent space as linear combinations of concept embeddings. By applying EDDP to the activations of a specific layer (the 12th residual block of an Xception network pre-trained on FaceForensics++), the authors jointly learn:

   • Decoding Directions ($W$): Vectors that act as concept detectors.
   • Encoding Directions ($S$): Vectors representing the specific directions in the latent space where concepts are encoded.
   • Thresholds ($b$): Decision boundaries for concept activation.

2. Visualization and Quantification:

   • Concept Presence Maps (CPMs): Binary maps generated by applying decoding directions to spatial patches, up-sampled to show where specific concepts are present.
   • Concept Contribution Maps (CCMs): Spatially aware heatmaps that quantify the influence of each concept on the final prediction logit, distinguishing between mere presence and causal contribution.

3. Validation and Analysis:

   • RCAV Sensitivity: Relative Concept Activation Vector scores are used to determine whether a concept biases the model toward "Real" or "Fake" classifications.
   • Semantic Mapping: Intersection over Union (IoU) scores map abstract concept directions to ground-truth facial regions (e.g., nose, eyes, mouth).
   • Faithfulness Assessment: The authors employ concept cloning (transplanting concept coefficients from a target sample to a source base) and targeted interventions (adding/removing concepts to flip predictions) to verify that the identified concepts are the primary drivers of the model's decision.

Key Contributions

• First Post-Hoc Concept-Based XAI for Deepfakes: This is the first application of post-hoc, concept-based XAI specifically within the deepfake detection domain, avoiding the need for model retraining or artifact-level supervision.
• Semantic Vocabulary Discovery: The work successfully uncovers a "semantic vocabulary" of 16 internal concepts learned by the detector, mapping abstract latent activations to human-interpretable features (e.g., "fake-mouth," "real-eyes," "fake-neck").
• Global and Local Understanding: The framework enables global model understanding by identifying which concepts drive predictions, while CCMs provide spatially aware, concept-based local explanations that pinpoint specific manipulation artifacts.
• Counterfactual Analysis: The method facilitates "what-if" analysis, demonstrating causal relationships between specific concepts and model outputs through systematic intervention.

Results

• Concept Identification: The analysis revealed distinct concepts that consistently bias the model toward "Fake" (e.g., $c_1, c_3, c_9, c_{10}, c_{12}, c_{14}$) and others toward "Real" (e.g., $c_0, c_5, c_7, c_8, c_{11}$). Some concepts were found to be non-discriminative.
• Semantic Grounding: Quantitative mapping via IoU confirmed that concepts align with specific facial regions (e.g., $c_5$ correlates with the nose, $c_{12}$ with fake eyes).
• Faithfulness: The concept cloning experiment achieved 87.34% accuracy in reproducing target predictions, and targeted interventions on misclassified samples achieved a 99.8% success rate in correcting the model's output. This confirms that the learned concept coefficients capture the essential features used for classification.
• Manipulation Specificity: The study observed that certain concepts exhibit high specificity to particular deepfake generation techniques (e.g., $c_3$ and $c_9$ are highly prevalent in DeepFakes and Face2Face), indicating the detector learns artifacts inherent to specific generative processes.

Significance and Claims
The paper claims that this approach provides a deeper comprehension of deepfake detection strategies by moving beyond binary predictions to reveal the intrinsic features learned during training. By grounding classification logic in identifiable facial regions and providing a mechanism to steer model predictions, the work enables more transparent and trustworthy detection systems. The authors emphasize that their method offers tight alignment between explanation and decision function, improving faithfulness without the operational overhead of retraining or architectural constraints.

Limitations
The authors modestly acknowledge several constraints:

• While the detector itself is not retrained, the EDDP model must still be learned on the specific representation space of the target model.
• The number of concepts is a manually defined hyperparameter, which risks redundancy or insufficient signal if not selected correctly.
• The identified concepts are optimized for the specific model architecture (Xception) and dataset (FaceForensics++) used; they may not be directly transferable to other models or datasets without a separate training process.