TL;DR
This paper introduces CorrRISE, a novel model-agnostic method for generating visual saliency maps that explain face verification decisions. Unlike previous methods, CorrRISE highlights both similar and dissimilar regions between face pairs, providing a more comprehensive explanation, and proposes a new quantitative evaluation framework for such methods.
What Problem Does This Paper Address?
Deep learning models have achieved remarkable accuracy in face recognition (FR), but their "black-box" nature makes it difficult to understand why they make certain decisions. This lack of interpretability is a significant barrier to trust and deployment, especially in sensitive applications like surveillance and access control. While visual saliency maps have emerged as a way to explain FR systems, existing methods often focus on classification tasks and struggle to provide meaningful explanations for face verification, which inherently involves comparing two images. Specifically, they often fail to clearly indicate why two faces are deemed a match or a non-match, particularly in the latter case. This paper aims to bridge this gap by developing an explainable face verification framework that provides insightful visual explanations for both matching and non-matching scenarios.
Methods
The core contribution is the CorrRISE (Correlation-based Randomized Input Sampling for Explanation) algorithm. CorrRISE is a model-agnostic approach, meaning it can be applied to any pre-trained face verification model without needing access to its internal architecture or requiring retraining.
The method works by:
- Mask Generation: Randomly generating a set of
Nbinary masks. These masks are applied to the input face images. - Perturbation and Score Calculation: For each mask, the original face images are perturbed (masked), and the perturbed images are fed into the face verification model. The model's output (e.g., similarity score between the two perturbed faces) is recorded. This process is repeated for all
Nmasks. - Correlation-based Saliency Map Generation: The key innovation lies in calculating the Pearson correlation between the recorded similarity scores and the generated masks, on a pixel-wise basis. This correlation reveals how much each pixel's presence or absence influences the model's decision.
- Similar Regions: Pixels with a positive correlation are highlighted as contributing to the similarity between the faces.
- Dissimilar Regions: Pixels with a negative correlation are highlighted as contributing to the dissimilarity.
This approach directly addresses the limitations of prior methods by disentangling similar and dissimilar regions, providing a more complete picture of the model's reasoning.
Additionally, the paper introduces a new quantitative evaluation methodology for saliency-based explanation methods in face verification. This methodology comprises two metrics:
- Deletion Metric: Measures how quickly the face verification accuracy drops when salient pixels (identified by the explanation method) are removed from the input images. A lower score indicates a more accurate explanation.
- Insertion Metric: Measures how quickly the face verification accuracy increases when salient pixels are inserted into a plain image. A higher score indicates a more accurate explanation.
Datasets
The paper uses the following datasets:
- MS1Mv2: Used for training the face recognition models.
- LFW (Labeled Faces in the Wild): Used for visual presentation and quantitative evaluation.
- CPLFW (Cross-pose Labeled Faces in the Wild): Used for visual presentation and quantitative evaluation, focusing on pose variations.
- LFR (LFR Face Dataset): Used for visual presentation.
- Webface-Occ (Webface-Occlusion Dataset): Used for visual presentation, focusing on occluded faces.
Key Insights and Ideas
- Disentangling Similar and Dissimilar Regions: The core insight is that a comprehensive explanation for face verification requires identifying both what makes two faces similar and what makes them dissimilar. Previous methods often only focused on one aspect, leading to incomplete explanations.
- Model-Agnosticism is Crucial: The ability to explain existing, deployed face verification systems without retraining them is highly valuable. CorrRISE achieves this, making it practical for real-world applications.
- Correlation as a Measure of Influence: Using Pearson correlation between mask perturbations and similarity scores is an effective way to quantify the influence of individual pixels on the model's decision.
- Need for Objective Evaluation: The paper highlights the long-standing issue of the lack of rigorous, objective evaluation metrics for saliency-based explanations, which has historically relied on subjective human evaluation. The proposed Deletion and Insertion metrics offer a more standardized way to compare methods.
- Sanity Check for Explanations: The paper demonstrates the importance of sanity checks to ensure that saliency maps are genuinely reflecting the model's decision-making process and not just producing generic heatmaps (e.g., focusing on the center of faces) regardless of the model's parameters.
Key Results
- Visual Explanations: CorrRISE generates visually compelling saliency maps that accurately highlight regions contributing to matching (similar regions) and non-matching (dissimilar regions) decisions across various scenarios, including standard verification, partially occluded faces, similar identities, and head pose variations.
- Quantitative Performance: In quantitative evaluations using the proposed Deletion and Insertion metrics on LFW, CPLFW, and Webface-Occ datasets, CorrRISE consistently outperforms adapted XAI methods like Grad-CAM and RISE. It also achieves better quantitative results than other specialized XFR explanation methods like xFace and MinPlus, despite them producing visually comparable results in some cases.
- Model-Agnostic Validation: Experiments with different state-of-the-art face recognition models (ArcFace, MagFace, AdaFace) show that CorrRISE's performance metrics remain consistent, validating its model-agnostic nature.
Limitations and Future Work
- Limitations: The paper acknowledges that while CorrRISE produces more stable saliency maps than many other approaches, some methods like xFace and MinPlus achieve comparable results. The paper also notes that some challenging cases, like very similar identities, can still lead to incorrect predictions by the FR model, and explanations for these failures are provided.
- Future Work: The paper's contribution of a new evaluation methodology and the CorrRISE algorithm itself opens avenues for future research in developing more robust and reliable explainable AI for face verification. The authors suggest that their work provides a foundation for further advancements in this area.
Summary updated