Introduction
The rise of Prediction-as-a-Service (PaaS) has created a fundamental tension between model confidentiality and data privacy. To address this, researchers developed secure inference protocols. The most efficient are "hybrid" protocols that combine Homomorphic Encryption (HE) for linear operations (like convolutions) and Multi-Party Computation (MPC) for non-linear steps (like ReLU activations).
However, these protocols typically assume "semi-honest" behavior, meaning they assume all parties follow the rules. This paper identifies a significant loophole: a malicious client can secretly "malleate" (shift) intermediate features during the MPC phase. Because the protocol cannot detect these shifts, the client can use them as a side-channel to reverse-engineer the model's weights and biases. The authors introduce SEEK (Safe-Error Extraction attacK) to exploit this vulnerability.
Methodology
SEEK treats model extraction as a "fault injection" attack. The core idea is to find critical points—specific input values where a tiny change in an internal feature causes the final classification label to flip.
Key Technical Components
- Safe-Error Strategy: By adding a large negative shift to specific features, the attacker can "zero out" parts of the network. This isolates and allows for the measurement of a single weight or bias without interference from the rest of the layer.
- Bisection Search: An algorithm is used to efficiently find the boundary between two classes by varying a single internal shift.
- Layer Independence: Unlike previous attacks that become more difficult as models get deeper, SEEK extracts each layer independently. This ensures the extraction process remains effective regardless of whether the model has 5 or 500 layers.
- Handling Complex Layers: The attack includes specific algorithms for dealing with standalone ReLU layers, combined Maxpool-ReLU layers, and the final fully-connected layer.
Experimental Evaluation
The researchers evaluated SEEK using a ResNet-18 model (approximately 11.7 million parameters) pre-trained on the ImageNet classification task. ImageNet is a standard large-scale benchmark, making it a realistic target for demonstrating the attack's potency.
Key Results
- Efficiency: On average, SEEK requires fewer than 50 queries to extract a single parameter.
- High Fidelity: The average relative error for extracted biases was roughly 0.0006%, and for weights, it was 0.004%.
- Label-Only Success: SEEK is significant because it works even when the service only returns the final class label (e.g., "Cat" vs "Dog"), which is the most restrictive and supposedly secure setup.
Insights and Future Work
The "aha moment" of SEEK is realizing that in hybrid protocols, the client holds "secret shares" of intermediate data. Since the server cannot see these shares, the client can add arbitrary values to them, turning a "secure" protocol into a playground for the attacker. Because SEEK tests whether a shift affects the criticality of the input rather than solving a global optimization problem, it avoids the numerical errors that plague other extraction attacks on deep networks.
While highly effective against semi-honest protocols, the authors note that moving to malicious-secure protocols (using cryptographic proofs) would mitigate this attack, though with significant computational overhead. Other potential defenses include shuffling features or using constant-round inference techniques like garbled circuits to make malleation harder for an adversary.
Summary updated