Abstract
Deep learning models are increasingly deployed in safety-critical domains, yet traditional verification methods focus on pixel-level perturbations that lack human interpretability. We present AuditAI, a framework that bridges classifiers with generative models to enable verification against semantic specifications. By utilizing a shared latent space, AuditAI allows auditors to define "unit tests" for human-understandable variations—such as changes in pose, lighting, or medical imaging conditions. Our approach provides formal safety guarantees, tighter verification bounds, and significantly improved robustness to domain shifts compared to standard pixel-based methods.
Introduction
As deep learning models move into healthcare and autonomous systems, the need for rigorous auditing becomes paramount. Traditional formal verification proves a model's robustness against $L_p$ norm perturbations—essentially ensuring that tiny, often invisible changes to pixels do not flip a classification. However, these mathematical proofs rarely align with real-world concerns. A radiologist is less concerned with adversarial noise and more concerned with whether a model remains accurate across different X-ray machine angles or patient orientations.
There exists a fundamental gap between the scalability of pixel-based verification and the interpretability required by human auditors. AuditAI addresses this by moving the verification target from the pixel space to a semantically meaningful latent space.
The AuditAI Framework
AuditAI creates a functional "bridge" between a standard classifier and a generative model, such as a Variational Autoencoder (VAE) or a Generative Adversarial Network (GAN). This architecture allows the system to translate abstract mathematical bounds into realistic image variations.
The framework operates through four primary components:
- Shared Latent Space: The classifier and generative model share a common encoder. This ensures the latent space—the compressed representation of the data—is semantically aligned. Specific dimensions in this space can be mapped to features like "pose," "lighting," or "expression."
- Semantic Unit Tests: Auditors define safety requirements as unit tests within this latent space (e.g., "Maintain >95% accuracy for all face rotations within 30 degrees").
- Latent Interval Bound Propagation (IBP): AuditAI adapts IBP to propagate bounds through the latent dimensions rather than raw pixels. By bounding a semantic dimension, the system mathematically proves the model's output remains stable across an entire range of realistic variation.
- Certified Training: These latent bounds are integrated into the loss function during training, forcing the model to be robust to specified semantic changes by design.
Certified Training and Latent Verification
By training models to be robust within the latent space, AuditAI achieves "tighter" bounds than pixel-space methods. Because the latent space is lower-dimensional and more structured, the mathematical propagation of uncertainty is less conservative, allowing the system to certify much larger ranges of variation that would be impossible to prove at the pixel level.
Experimental Evaluation
The framework was evaluated across diverse domains, including ImageNet (general objects), CheXpert and NIH (medical X-rays), LSUN (architectural towers), and FFHQ (high-quality faces).
Key Results
- Superior Scaling: On ImageNet, AuditAI-trained models tolerated 25% larger variations than standard pixel-based certified training while maintaining a high verified accuracy of 88%.
- Generalization to Domain Shift: In medical imaging, models trained on the CheXpert dataset using AuditAI outperformed pixel-based approaches by 5% when deployed on the NIH dataset. This suggests that semantic auditing encourages the model to ignore irrelevant "noise" and focus on core diagnostic features.
- Human-Validated Realism: To ensure the "semantic variations" were realistic, the authors conducted a study where humans found generated images indistinguishable from real ones up to a perturbation threshold of $\epsilon \leq 1.0$.
Discussion and Future Work
AuditAI demonstrates that verifying models in a semantically aligned latent space offers a path toward more trustworthy AI. However, the framework's efficacy is currently tied to the quality of the underlying generative model; if a GAN cannot represent a specific rare feature, the auditor cannot test for it.
Future research will focus on integrating human domain experts, such as radiologists, directly into the loop to define critical semantic dimensions. While currently focused on computer vision, the principles of AuditAI are extensible to language modeling and other complex decision-making applications where human-interpretable safety guarantees are essential.
Summary updated