Introduction
As deep learning models scale, "split inference" has become a standard for deploying AI on resource-constrained edge devices. By partitioning a Deep Neural Network (DNN), the first few layers run locally on the edge (e.g., an IoT sensor), while the remaining layers execute on a powerful cloud server. While this distribution optimizes performance, it introduces a critical security vulnerability: the communication link between the edge and the cloud. This research demonstrates that an attacker intercepting these intermediate features can launch highly effective, transferable adversarial attacks, even against black-box target models.
The Vulnerability of Split Inference
Traditional adversarial research focuses on "single-sided" models where the attacker only interacts with the final output. In contrast, distributed models expose "intermediate features"—the internal representations of data as it passes through the network.
If an attacker can sniff these features, they gain a privileged view of the model’s internal logic. This allows them to build a "surrogate" model that mimics the target's behavior far more accurately than a model trained only on final classification labels.
Methodology: Exploiting the Wire
The proposed attack strategy involves two primary phases:
1. Shape Reconstruction
Intermediate data is often transmitted as flattened vectors to minimize protocol overhead, hiding the original tensor dimensions (e.g., height, width, channels). The researchers utilize statistical covariance analysis and autocorrelation on intercepted batches to identify spatial patterns. Because spatial correlation in images survives flattening, the covariance matrix reveals a grid-like structure that allows the attacker to accurately reconstruct the original feature map dimensions.
2. Feature Distillation with Adaptation
With the reconstructed shape, the attacker trains a surrogate model. To bridge the gap between the surrogate's architecture and the unknown target architecture, an Adaptation Module (comprising lightweight encoder/decoder layers) is used. This module aligns the surrogate's internal features with the intercepted ones. The surrogate is trained using a combined loss function that matches both the target's final outputs and its internal representations, forcing the surrogate to "think" like the target at a structural level.
Key Insights and Results
The study, conducted using the CIFAR-10 benchmark, reveals several critical findings:
- Enhanced Transferability: Adversarial examples generated on the feature-informed surrogate transferred to the target model with significantly higher success rates. In white-box transfer scenarios, success increased by up to 51%.
- Black-Box Dominance: In query-based attacks (GFCS) against a ResNet56 target, the success rate jumped from 69% (standard label-only attack) to 96% when leveraging intermediate features.
- Efficiency and Stealth: These attacks require fewer queries to the target model, making them harder to detect. Furthermore, the resulting adversarial perturbations are smaller and less visible to the human eye.
- Accuracy as a Proxy: A key discovery is the strong correlation between a surrogate's clean accuracy and its attack transferability. The surrogate that performs best on normal data is typically the most effective at generating adversarial attacks.
Conclusion and Future Directions
This research serves as a "call to arms" for the AI security community. While split inference offers operational benefits, it creates a significant attack surface. Current defenses must evolve beyond simple encryption to include feature obfuscation—techniques that allow the cloud to process data while preventing eavesdroppers from reconstructing the model's internal logic. Future work will explore whether these statistical vulnerabilities extend to other architectures, such as Transformers or Recurrent Neural Networks.
Summary updated