Trojan attacks on deep neural networks (DNNs) exploit a backdoor embedded in a DNN model that can hijack any input with an attacker’s chosen signature trigger. Emerging defence mechanisms are mainly designed and validated on vision domain tasks (e.g., image classification) on 2D Convolutional Neural Network (CNN) model architectures; a defence mechanism that is general across vision, text, and audio domain tasks is demanded. This work designs and evaluates a run-time Trojan detection method exploiting STR ong I ntentional P erturbation of inputs that is a multi-domain input-agnostic Trojan detection defence across Vi sion, T ext and A udio domains—thus termed as STRIP-ViTA. Specifically, STRIP-ViTA is demonstratively independent of not only task domain but also model architectures. Most importantly, unlike other detection mechanisms, it requires neither machine learning expertise nor expensive computational resource, which are the reason behind DNN model outsourcing scenario—one main attack surface of Trojan attack. We have extensively evaluated the performance of STRIP-ViTA over: i) CIFAR10 and GTSRB datasets using 2D CNNs for vision tasks; ii) IMDB and consumer complaint datasets using both LSTM and 1D CNNs for text tasks; and iii) speech command dataset using both 1D CNNs and 2D CNNs for audio tasks. Experimental results based on more than 30 tested Trojaned models (including publicly Trojaned model) corroborate that STRIP-ViTA performs well across all nine architectures and five datasets. Overall, STRIP-ViTA can effectively detect trigger inputs with small false acceptance rate (FAR) with an acceptable preset false rejection rate (FRR). In particular, for vision tasks, we can always achieve a 0 percent FRR and FAR given strong attack success rate always preferred by the attacker. By setting FRR to be 3 percent, average FAR of 1.1 and 3.55 percent are achieved for text and audio tasks, respectively. Moreover, we have evaluated STRIP-ViTA against a number of advanced backdoor attacks and compare its effectiveness with other recent state-of-the-arts.

Anomaly detection has been known as an effective technique to detect faults or cyber-attacks in industrial control systems (ICS). Therefore, many anomaly detection models have been proposed for ICS. However, most models have been implemented and evaluated under specific circumstances, which leads to confusion about choosing the best model in a real-world situation. In other words, there still needs to be a comprehensive comparison of state-of-the-art anomaly detection models with common experimental configurations. To address this problem, we conduct a comparative study of five representative time series anomaly detection models: InterFusion, RANSynCoder, GDN, LSTM-ED, and USAD. We specifically compare the performance analysis of the models in detection accuracy, training, and testing times with two publicly available datasets: SWaT and HAI. The experimental results show that the best model results are inconsistent with the datasets. For SWaT, InterFusion achieves the highest $F1$-$score$ of 90.7% while RANSynCoder achieves the highest $F1$-$score$ of 82.9% for HAI. We also investigate the effects of the training set size on the performance of anomaly detection models. We found that about 40% of the entire training set would be sufficient to build a model producing a similar performance compared to using the entire training set.

The Internet of Things (IoT) has seen remarkable advancements in recent years, leading to a paradigm shift in the digital landscape. However, these technological strides have introduced new challenges, particularly in cybersecurity. IoT devices, inherently connected to the internet, are susceptible to various forms of attacks. Moreover, IoT services often handle sensitive user data, which could be exploited by malicious actors or unauthorized service providers. As IoT ecosystems expand, the convergence of traditional and cloud-based systems presents unique security threats in the absence of uniform regulations. Cloud-based IoT systems, enabled by Platform-as-a-Service (PaaS) and Infrastructure-as-a-Service (IaaS) models, offer flexibility and scalability but also pose additional security risks. The intricate interaction between these systems and traditional IoT devices demands comprehensive strategies to protect data integrity and user privacy. This paper highlights the pressing security concerns associated with the widespread adoption of IoT devices and services. We propose viable solutions to bridge the existing security gaps while anticipating and preparing for future challenges. This paper provides a detailed survey of the key security challenges that IoT services are currently facing. We also suggest proactive strategies to mitigate these risks, thereby strengthening the overall security of IoT devices and services.

Extensive use of unmanned aerial vehicles (commonly referred to as a “drone”) has posed security and safety challenges. To mitigate security threats caused by flights of unauthorized drones, we present a framework called SENTINEL (Secure and Efficient autheNTIcation for uNmanned aErial vehicLes) under the Internet of Drones (IoD) infrastructure. SENTINEL is specifically designed to minimize the computational and traffic overheads caused by certificate exchanges and asymmetric cryptography computations that are typically required for authentication protocols. SENTINEL initially generates a flight session key for a drone having a flight plan and registers the flight session key and its flight plan into a centralized database that can be accessed by ground stations. The registered flight session key is then used as the message authentication code key to authenticate the drone by any ground station while the drone is flying. To demonstrate the feasibility of the proposed scheme, we implemented a prototype of SENTINEL with ECDSA, PBKDF2 and HMAC-SHA256. The experiment results demonstrated that the average execution time of the authentication protocol in SENTINEL was about 3.1 times faster than the “TLS for IoT” protocol. We also formally proved the security of SENTINEL using ProVerif that is an automatic cryptographic protocol verifier.

Federated learning (FL) and split learning (SL) are state-of-the-art distributed machine learning techniques to enable machine learning training without accessing raw data on clients or end devices. However, their comparative training performance under real-world resource-restricted Internet of Things (IoT) device settings remains barely studied. This work provides empirical comparisons of FL and SL in real-world IoT settings regarding (i) learning performance with heterogeneous data distributions and (ii) on-device execution overhead. Our analyses in this work demonstrate that the learning performance of SL is better than FL under an imbalanced data distribution but worse than FL under an extreme non-IID data distribution. Recently, FL and SL are combined to form splitfed learning (SFL) to leverage each of their benefits (e.g., parallel training of FL and lightweight on-device computation requirement of SL). Our work considers FL, SL, and SFL, and mounts them on Raspberry Pi devices to evaluate their performance, including training time, communication overhead, power consumption, and memory usage with resource-restricted IoT devices. Besides evaluations, we apply two optimizations. First, we generalize SFL by carefully examining the possibility of a hybrid type of model training at the server-side. The generalized SFL merges sequential (dependent) and parallel (independent) processes of model training and thus is beneficial to a system with a large scale of IoT devices, specifically at the server-side operations. Second, we propose pragmatic techniques to substantially reduce the communication overhead by up to four times for the SL and (generalized) SFL.