Secure On-Device Video OOD Detection Without Backpropagation

TL;DR

This paper proposes SecDOOD, a cloud-device collaboration framework for secure, backpropagation-free on-device OOD detection. It combines a HyperNetwork-based personalized parameter generation module trained in the cloud with a dynamic feature sampling and encryption strategy that protects privacy by encrypting only the most informative feature channels. The device performs lightweight inference with cloud-derived personalized parameters, achieving performance comparable to fully fine-tuned models while reducing on-device computation by a factor of about 3 and maintaining privacy through selective encryption. Extensive experiments across five datasets and near-/far-OOD settings demonstrate SecDOOD's robustness, efficiency, and adaptability to diverse deployment scenarios, highlighting its practical impact for edge AI in safety-critical applications.

Abstract

Out-of-Distribution (OOD) detection is critical for ensuring the reliability of machine learning models in safety-critical applications such as autonomous driving and medical diagnosis. While deploying personalized OOD detection directly on edge devices is desirable, it remains challenging due to large model sizes and the computational infeasibility of on-device training. Federated learning partially addresses this but still requires gradient computation and backpropagation, exceeding the capabilities of many edge devices. To overcome these challenges, we propose SecDOOD, a secure cloud-device collaboration framework for efficient on-device OOD detection without requiring device-side backpropagation. SecDOOD utilizes cloud resources for model training while ensuring user data privacy by retaining sensitive information on-device. Central to SecDOOD is a HyperNetwork-based personalized parameter generation module, which adapts cloud-trained models to device-specific distributions by dynamically generating local weight adjustments, effectively combining central and local information without local fine-tuning. Additionally, our dynamic feature sampling and encryption strategy selectively encrypts only the most informative feature channels, largely reducing encryption overhead without compromising detection performance. Extensive experiments across multiple datasets and OOD scenarios demonstrate that SecDOOD achieves performance comparable to fully fine-tuned models, enabling secure, efficient, and personalized OOD detection on resource-limited edge devices. To enhance accessibility and reproducibility, our code is publicly available at https://github.com/Dystopians/SecDOOD.

Figures (4)

• Figure 1: Comparison of traditional OOD detection methods and our proposed SecDOOD. Conventional approaches require training or fine-tuning the classifier/backbone model for OOD detection, which is computationally demanding and challenging to execute on resource-constrained devices. In contrast, SecDOOD introduces a hypernetwork-based solution that eliminates the need for backpropagation on the device, enabling efficient and lightweight OOD detection.
• Figure 2: Overview of the proposed SecDOOD framework. SecDOOD leverages cloud-device collaboration to enable efficient and privacy-preserving OOD detection on edge devices. During deployment, the edge device extracts feature representations from incoming data and applies dynamic feature sampling to select the most informative channels. These selected features are securely encrypted and transmitted to the cloud, where a HyperNetwork-based module generates personalized model parameters tailored to the device-specific data distribution. The personalized parameters are then sent back to the device and encrypted, where they are decrypted and used for inference. This approach ensures adaptability to user-specific distributions while maintaining computational efficiency and data privacy.
• Figure 3: Encryption and decryption time per sample under different feature channel encryption ratios. Encrypting all channels (100%) results in significantly higher time costs, while encrypting 50% and 25% of the channels shows similar computational efficiency. Based on this analysis, we select 50% encryption as the optimal balance between security and efficiency.
• Figure 4: FLOPs comparison between traditional on-device training and SecDOOD. Traditional methods require full model training on the device, incurring high computational costs. SecDOOD offloads training to the cloud, reducing on-device computation to feature extraction and inference, achieving a 3× efficiency gain.