TL;DR
Federated Learning (FL) is often touted as a privacy-preserving savior for data-scarce fields like hardware assurance. This paper shatters that assumption by introducing a data-free Membership Inference Attack (MIA) that uses publicly available standard cell libraries to reconstruct private chip images from model updates. The attack successfully leaks highly sensitive intellectual property, including specific circuit layers and manufacturing technology nodes, without requiring any private auxiliary datasets.
What Problem Does This Paper Address?
In the semiconductor industry, hardware assurance relies heavily on deep learning to detect Trojans, reverse-engineer chips, and analyze defects from Scanning Electron Microscopy (SEM) images. However, acquiring this data requires expensive, destructive de-layering processes, leading to severe data scarcity.
Federated Learning (FL) allows multiple organizations to collaboratively train a shared model without directly exchanging their sensitive raw datasets. While FL protects raw data from direct exposure, it remains vulnerable to Membership Inference Attacks (MIAs)—where an adversary tries to determine if a specific data record was used during training. Traditional MIAs rely on "shadow datasets" that statistically mirror the target data. In hardware assurance, this assumption fails because proprietary chip layouts and SEM images are highly guarded secrets that vary drastically across different manufacturing processes. This paper addresses this gap by developing an MIA that requires zero private auxiliary data.
Methods
The authors propose a three-stage pipeline operating under an "honest-but-curious" threat model, where a passive adversarial global server observes model updates without altering the training protocol.
- Federated Learning: Clients collaboratively train a U-Net image segmentation model using the standard Federated Averaging (FedAvg) algorithm.
- Guided Gradient Inversion Attack (GIA): The adversarial server intercepts a client's model updates and extracts the gradients. Instead of performing a standard blind image reconstruction, the attacker leverages publicly available Standard Cell Library Layouts (SCLLs) as fixed "dummy labels" to guide the inversion process. Crucially, they introduce a composite loss function featuring a negative forward pass loss term ($-\lambda_{dummy}L_{dummy}$). This term penalizes the optimizer for simply memorizing or fitting the dummy label, forcing it to prioritize matching the true gradient signal.
- Membership Inference via Mean Threshold: The raw reconstructed images are post-processed (using Gaussian blur, Otsu's thresholding, and morphological operations) into clean binary masks. The attack calculates the Dice similarity coefficient between these masks and the guiding SCLLs. Finally, a global mean threshold classifies whether the data point belongs to a member or a non-member.
Datasets
The evaluation utilizes Standard Cell Library Layouts (SCLLs) from Synopsys’ Open Educational Design Kit (SAED) for both 32nm and 90nm technology nodes. The authors filtered the libraries to isolate core logic gates (AND, NAND, OR, etc.). To simulate realistic hardware assurance tasks, they used the REFICS tool to generate a synthetic dataset of 141 SEM images and masks embedded with realistic shot noise and dwelling time parameters.
Key Insights and Ideas
- Public Libraries as Priors: Attackers do not need clones of proprietary datasets to break FL privacy. Publicly available Process Design Kits (PDKs) and standard cell libraries provide enough structural prior information to successfully guide gradient inversion attacks.
- The Post-Processing Bottleneck: Complex, fine-grained hardware structures (like metal layers or smaller 32nm nodes) are highly sensitive to standard image cleanup pipelines. This creates an evaluation bottleneck where high-quality raw reconstructions get degraded during post-processing, artificially lowering similarity scores and masking the true extent of data leakage.
- Amplifying the Gap via Negative Loss: By subtracting the forward pass loss during optimization, the GIA intentionally and disproportionately degrades non-member reconstructions while keeping member reconstructions relatively intact. This counter-intuitive trick widens the measurable quality gap, successfully overcoming the post-processing bottleneck for complex data.
Key Results
- Inter-Layer Inference (Metal vs. Diffusion): When distinguishing between circuit layers using standard GIA ($\lambda_{dummy}=0$), the attack achieved a near-perfect Area Under the Curve (AUC) of 0.9804 for simpler diffusion layers, but dropped to 0.8868 for complex metal layers due to the post-processing bottleneck. Activating the negative loss term ($\lambda_{dummy}=5$) successfully amplified the quality gap, boosting the metal layer attack AUC to 0.9752.
- Intra-Layer Inference (32nm vs. 90nm): Distinguishing between different technology nodes of the same layer is exceptionally difficult due to structural similarities. At baseline ($\lambda_{dummy}=0$), the attack on the 32nm node scored a poor 0.6688 AUC. However, applying the optimized negative loss term ($\lambda_{dummy}=5$) skyrocketed the attack performance to a near-perfect 0.9916 AUC.
Limitations and Future Work
- Scalability: The experiments were conducted in a controlled, two-client setting. Future work must scale the simulation to include more clients and evaluate performance on larger, more complex neural network architectures.
- Hyperparameter Automation: The optimal weight for the negative loss term ($\lambda_{dummy}$) is highly data-dependent and currently requires manual tuning. Developing an automated method to tune this hyperparameter dynamically would make the attack far more generalizable.
- Defense Evaluation: While this work highlights a severe vulnerability, a critical next step is evaluating how robust common privacy-preserving defenses (such as Differential Privacy or cryptographic aggregation) are against this style of guided attack.
Summary updated