The Vulnerability of Collaborative Inference
Deep learning models are growing exponentially in size, making them difficult to deploy on resource-constrained IoT devices like smart cameras or wearables. To bridge this gap, the industry often relies on collaborative inference: an edge device computes the first few layers of a neural network and transmits the intermediate "representations" to a powerful cloud server to complete the task.
While raw data never leaves the device, this process is not inherently private. Recent "Model Inversion" attacks demonstrate that a malicious server can reconstruct the original input (e.g., a user's face) from these intermediate features. Furthermore, "Model Completion" attacks allow the server to intercept the final prediction before it reaches the user. Existing defenses, such as Differential Privacy (DP) or heavy data compression, act like a sledgehammer—they protect privacy but severely degrade the model's accuracy.
Introducing InfoScissors
To address these vulnerabilities, researchers introduce InfoScissors, a novel training strategy designed to minimize the Mutual Information (MI) between the model’s intermediate outputs and two sensitive targets: the raw input ($x$) and the final label ($y$).
The core innovation of InfoScissors lies in how it calculates and minimizes MI. Most prior works rely on the Variational Information Bottleneck (VIB). However, the authors mathematically demonstrate that VIB is too aggressive for collaborative inference. It attempts to force intermediate data into a simple Gaussian distribution, which often indiscriminately deletes the very features the cloud needs to complete its task.
Instead, InfoScissors leverages the Contrastive Log-ratio Upper Bound (CLUB). This approach is far more "surgical," focusing on the conditional distribution to filter out reconstruction-specific information while preserving task-specific features. The defense operates on three fronts:
- Input Protection: It regularizes the device-side model to ensure the intermediate representation contains minimal information about the raw input.
- Prediction Protection: It regularizes the server-side features to ensure they do not leak the final label.
- Adversarial Training: During training, the system uses an auxiliary "generator" (attempting to reconstruct the image) and an auxiliary "classifier" (attempting to guess the label). The main model learns to defeat these internal attackers, effectively hiding private details.
Importantly, this defense is "concealed." The cloud server does not need to change its architecture or even know the defense is active; the privacy-preserving filtering is baked directly into the model weights during the collaborative training phase.
Key Results and Performance
The researchers evaluated InfoScissors using standard computer vision benchmarks (CIFAR-10 and CIFAR-100) with a ResNet18 backbone. They simulated a "worst-case" privacy scenario where the edge device holds only the very first convolutional layer, maximizing potential leakage.
InfoScissors significantly outperformed five state-of-the-art baselines, including DP and Mutual Information Defense (MID), in the privacy-utility trade-off.
- Input Privacy: On CIFAR-10, InfoScissors reduced the Structural Similarity (SSIM) of reconstructed images to below 0.2—rendering them completely unrecognizable—with less than a 2% drop in primary model accuracy. In contrast, baseline methods typically required a 10-15% accuracy sacrifice to achieve similar privacy levels.
- Prediction Privacy: The defense successfully reduced a malicious server's "attack accuracy" to roughly 10% (equivalent to a random guess) while sacrificing less than 0.5% of the primary model's accuracy.
- Robustness: InfoScissors maintained its defensive capabilities even against "white-box" attacks (where the attacker knows the model's structure) and "adaptive" attacks tailored to circumvent the defense.
Limitations and Future Directions
While highly effective, the authors acknowledge a few limitations. The current study focuses on a one-to-one edge-cloud relationship, whereas real-world deployments often involve many devices collaborating with a single server. Additionally, while the defense adds no latency during inference, the training process is computationally heavier due to the auxiliary networks. Finally, future work will need to explore how this defense scales as different split points within the network architecture are utilized.
Summary updated