RobustEdge: Low Power Adversarial Detection for Cloud-Edge Systems

Abhishek Moitra, Abhiroop Bhattacharjee, Youngeun Kim, Priyadarshini Panda

Abstract

In practical cloud-edge scenarios, where a resource-constrained edge performs data acquisition and a cloud system performs inference tasks with a deep neural network (DNN), adversarial robustness is critical for reliability. Adversarial detection is a prime defense technique, but traditional detectors are often attached to the classifier in the cloud, requiring high computational overhead and causing energy wastage by transmitting malicious data. RobustEdge proposes Quantization-enabled Energy Separation (QES) training with "early detection and exit" to perform low-cost adversarial detection directly at the edge. By blocking adversarial data transmission, RobustEdge improves both the adversarial robustness and energy efficiency of the Cloud-Edge system.

The Problem: Energy-Draining Adversarial Attacks

Deep Neural Networks (DNNs) are the backbone of modern cloud-edge systems, where low-power devices like smart cameras capture data and send it to powerful cloud servers. However, DNNs are notoriously vulnerable to adversarial attacks—subtle, intentional perturbations to input data that cause misclassification.

Currently, most adversarial detection happens in the cloud. This creates two major problems:

  1. Energy Wastage: The edge device spends significant energy transmitting malicious data that will eventually be discarded.
  2. Computational Overhead: Existing detectors are often as complex as the classifiers themselves, making them too heavy for resource-constrained edge hardware.
Comparison of standard cloud-edge systems (a), cloud-based adversarial detection (b), and the proposed RobustEdge edge-based detection (c).

Methods: QES and Early Exit

RobustEdge introduces a standalone, 3-layer convolutional detector designed for the edge, utilizing two primary innovations:

  • Quantization-enabled Energy Separation (QES): Instead of final labels, RobustEdge analyzes "energy signatures"—the magnitude of weighted summation outputs at each layer. Adversarial samples typically exhibit different energy distributions than natural ones. QES trains the detector to maximize this "energy gap" while quantizing the model (to 12-bit or 16-bit) for hardware efficiency.
  • Early Detection and Exit: To minimize power, the detector establishes "confidence boundaries" at each layer. If an input’s energy signature is clearly natural or adversarial at Layer 1, the device makes an immediate decision and "exits," saving computation cycles for subsequent layers.
Energy value distributions for natural and adversarial data at the first layer, illustrating the 'undecided region' used for early exit logic.

Key Results

The researchers evaluated RobustEdge using standard benchmarks like CIFAR10, CIFAR100, and TinyImagenet against a battery of gradient-based and black-box attacks. Key findings include:

  • High Accuracy: The 16-bit and 12-bit detectors achieved an Area Under the Curve (AUC) score > 0.9 across most gradient-based attacks.
  • Massive Energy Savings: Hardware evaluations on a 45nm CMOS digital accelerator showed that RobustEdge requires 25x lower energy for detection than prior state-of-the-art methods.
  • System Efficiency: By blocking 100% of adversarial transmissions, total system energy efficiency improved by > 166x.
  • Early Exit Success: The "early exit" strategy reduced the detector's own energy consumption by 66% without sacrificing significant accuracy.

Conclusion

RobustEdge demonstrates that "failing fast" at the edge is a superior strategy for cloud-edge security. By shifting detection to the point of data acquisition and utilizing energy-based signatures, the framework provides a scalable, classifier-agnostic solution that drastically reduces the energy footprint of secure IoT systems. Future work will focus on improving robustness against score-based attacks and adaptive adversaries who may attempt to mimic natural energy signatures.

Summary updated

Abstract

In practical cloud-edge scenarios, where a resource constrained edge performs data acquisition and a cloud system (having sufficient resources) performs inference tasks with a deep neural network (DNN), adversarial robustness is critical for reliability and ubiquitous deployment. Adversarial detection is a prime adversarial defence technique used in prior literature. However, in prior detection works, the detector is attached to the classifier model and both detector and classifier work in tandem to perform adversarial detection that requires a high computational overhead which is not available at the low-power edge. Therefore, prior works can only perform adversarial detection at the cloud and not at the edge. This means that in case of adversarial attacks, the unfavourable adversarial samples must be communicated to the cloud which leads to energy wastage at the edge device. Therefore, a low-power edge-friendly adversarial detection method is required to improve the energy efficiency of the edge and robustness of the cloud-based classifier. To this end, RobustEdge proposes Quantization-enabled Energy Separation (QES) training with "early detection and exit" to perform edge-based low cost adversarial detection. The QES-trained detector implemented at the edge blocks adversarial data transmission to the classifier model, thereby improving adversarial robustness and energy-efficiency of the Cloud-Edge system.
BibTeX
@article{moitra2024robustedge,
  author = {Abhishek Moitra and Abhiroop Bhattacharjee and Youngeun Kim and Priyadarshini Panda},
  title = {{RobustEdge: Low Power Adversarial Detection for Cloud-Edge Systems}},
  journal = {IEEE Trans. Emerg. Top. Comput. Intell.},
  year = {2024},
  url = {https://openreview.net/forum?id=S9MuH4SDON},
}
Figure