Introduction
As Deep Neural Networks (DNNs) become more complex and computationally expensive to train, the industry has shifted toward a model where pre-trained networks are treated as Intellectual Property (IP) cores. While this "IP-based" design flow accelerates development, it introduces significant security risks. Malicious actors in the supply chain can inject faults into model parameters—such as weights and biases—to create backdoors or cause specific misclassifications. Validating these "black-box" models is challenging because users often lack access to the internal architecture or the original training data.
The Validation Framework
To address these vulnerabilities, this paper proposes a practical validation scheme that allows users to verify the integrity of a DNN IP using a minimal set of functional tests. The framework relies on a cooperative model between the IP vendor and the user. The vendor generates a set of functional tests designed to maximize internal parameter activation and provides these alongside the model. The user then runs these tests to ensure the output matches the expected results provided by the vendor.
Methodology: Parameter Coverage
The core innovation of this approach is the shift from "neuron coverage" to Parameter Coverage. Traditional hardware testing often focuses on ensuring every neuron fires. However, a neuron can be active even if some of its incoming weights are tampered with, provided those specific weights do not significantly influence the final output for a given test case. A parameter is only considered "covered" if a perturbation in its value propagates to the network's output.
To maximize this coverage efficiently, the authors employ a two-stage strategy:
- Iterative Selection from Training Set: A greedy algorithm selects training samples that activate the highest number of previously un-activated parameters. Since DNNs are optimized for their training data, these samples are naturally effective at exercising the network's primary pathways.
- Gradient-based Test Generation: To reach "bottleneck" parameters that training data might miss, the authors generate synthetic test cases. Using gradient descent, they modify input pixels to minimize loss for a target class, effectively "forcing" the network to use specific internal pathways. These synthetic images often appear as exaggerated, ghostly versions of the target features, concentrating the information needed to trigger elusive neural structures.
Key Results
The researchers evaluated their method using MNIST and CIFAR-10 benchmarks. The "Combined Method"—which starts with training data selection and switches to synthetic generation—demonstrated superior efficiency. On the CIFAR-10 model, the method achieved 92% parameter coverage with only 30 test cases, significantly outperforming standard training sample selection.
In terms of security, the framework proved highly effective at detecting malicious tampering. On the MNIST dataset, using just 50 test cases, the proposed method achieved a 97.3% detection rate for Single Bias Attacks (SBA). This significantly outperformed traditional neuron coverage metrics, which caught only 89.1% of attacks under identical conditions.
Conclusion
This research demonstrates that high-integrity validation of DNN IPs is possible with a very small number of functional tests, provided those tests are designed to maximize parameter coverage rather than just neuron activity. While the current framework assumes the vendor provides the tests, it sets the stage for future work where users might independently generate validation suites to protect against potentially malicious vendors in a fully untrusted supply chain.
Summary updated