The ISC International Journal of Information Security

Abstract With the growing adoption of home robots, securing them via efficient intrusion detection systems (IDS) is increasingly vital.
To address the energy and computational constraints of robotic platforms, we adopt a distributed IDS architecture: a resource-rich remote component performs in-depth detection and analysis of operating system logs and network traffic using deep learning, while a lightweight component on the robot controller enables rapid detection.
Remotely estimated attack probabilities are fed into a risk calculator, which then guides a compact ensemble model running on the controller for real-time, risk-based detection. This approach is driven by the need to achieve higher detection performance on the resource-constrained controller.
Rather than relying solely on attack frequency, we enhance detection performance by incorporating self-attention mechanisms to capture dependencies and contextual relationships among attacks.
Evaluated on a custom home robot dataset, our approach demonstrates that modeling attack interdependencies significantly improves detection performance over frequency-only methods. The resulting system achieves high performance without compromising responsiveness, proving that attention-enhanced, risk-aware detection is both feasible and effective in resource-constrained robotic environments.

Abstract Today, with the advancement of science and technology, the use of smartphones has become very common, and the Android operating system has been able to gain lots of popularity in the meantime. However, these devices face many
security challenges, including malware. Malware may cause many problems in both the security and privacy of users. So far, the state-of-the-art method in malware detection is based on deep learning, however, this approach requires a lot of computing resources and leads to high battery usage, which is unacceptable in smartphone devices. This paper proposes the knowledge distillation approach for lightening android malware detection. To this end, first, a heavy model is taught and then with the knowledge distillation approach, its knowledge is transferred to a light model called student. To simplify the learning process, soft labels are used here. The resulting model, although slightly less accurate in identification, has a much smaller size than the heavier model. Moreover, ensemble learning was proposed to recover the dropped accuracy. We have tested the proposed approach on CISC datasets including dynamic and static features, and the results show that the proposed method is not only able to lighten the model up to 99%, but also maintain the accuracy of the lightened model to the extent of the heavy model.