Introduction
Deep Neural Networks (DNNs) have demonstrated remarkable performance across various domains, yet they remain notoriously fragile. Small, often imperceptible perturbations—known as adversarial examples—can lead to catastrophic failures in high-stakes environments like autonomous driving. While "robust training" techniques exist to mitigate these vulnerabilities, their adoption in industry has been hindered by significant engineering overhead. Most existing tools require models to be manually rewritten in specialized languages or restricted to a narrow set of supported layers.
To bridge this gap, we present PaRoT, a practical framework built on TensorFlow that automates robust training. By automatically transforming standard computation graphs, PaRoT makes formal verification techniques accessible for real-world, industrial-scale applications without requiring extensive code modifications.
Methodology: Automated Robust Training
The core of PaRoT is based on Differentiable Abstract Interpretation. Rather than training on discrete data points, the framework treats inputs as high-dimensional "regions" (e.g., boxes or zonotopes). These regions are propagated through the network to ensure that every possible point within the input region leads to a safe and correct output.
Key Technical Innovations
- Automatic Transformer Generation: PaRoT employs a graph-traversal algorithm to convert standard TensorFlow computation graphs into "abstract" versions. This allows developers to use existing models seamlessly.
- HybridZonotope Domains: The framework utilizes Hybrid Zonotopes, which offer a superior balance between the computational efficiency of simple bounding boxes and the precision of complex zonotopes, effectively tracking correlations between pixels.
- Refined Activation Transformers: The authors derived tighter mathematical bounds for non-linear activation functions like Sigmoid, significantly reducing the approximation errors common in formal verification.
Beyond standard $\epsilon$-ball perturbations, PaRoT supports custom properties such as Fourier attacks (modeling weather or lens grime) and brightness shifts, providing a more realistic assessment of sensor robustness.
Experimental Evaluation
We evaluated PaRoT across three primary benchmarks:
- MNIST & CIFAR10: Standard datasets for handwritten digit and object recognition.
- Traffic Light Detection: A real-world industrial dataset used to demonstrate the framework's utility in an autonomous driving context.
Results and Discussion
PaRoT achieves performance parity with previous state-of-the-art frameworks like DiffAI while offering vastly superior integration capabilities. On the MNIST dataset, a standard model was found to be 100% susceptible to certain adversarial attacks; however, a model trained with PaRoT reduced this susceptibility to approximately 30% while maintaining high standard accuracy.
Case Study: Traffic Light Detection
The framework's practical utility is best demonstrated in the traffic light detection case study. Standard models can be easily confused by perturbations that mimic real-world sensor noise. PaRoT-trained models successfully "fix" these vulnerabilities, correctly identifying signals even under adversarial conditions.
A key observation during testing was the inherent trade-off between standard accuracy and robustness. PaRoT allows developers to navigate this trade-off by adjusting a mixing factor ($\lambda$) within the loss function, enabling fine-tuned performance based on specific safety requirements.
Conclusion and Future Work
PaRoT shifts robust training from a research curiosity to a viable developer tool by automating the complex engineering tasks associated with formal verification. While the current implementation handles moderate-sized Convolutional Neural Networks (CNNs), scaling HybridZonotopes to massive, state-of-the-art architectures remains a computational challenge. Future work will focus on expanding the library of supported TensorFlow operations and exploring stochastic approximations to further accelerate training for very deep networks.
Summary updated