Search

The introduction of deep learning technology can improve the performance of cyber-physical systems (CPSs) in many ways. However, this also brings new security issues. To tackle these challenges, this paper explores the vulnerabilities of deep learning-based unmanned aerial vehicles (UAVs), which are typical CPSs. Although many research works have been reported previously on adversarial attacks of deep learning models, only few of them are concerned about safety-critical CPSs, especially regression models in such systems. In this paper, we analyze the problem of adversarial attacks against deep learning-based UAVs and propose two adversarial attack methods against regression models in UAVs. Experiments demonstrate that the proposed non-targeted and targeted attack methods both can craft imperceptible adversarial images and pose a considerable threat to the navigation and control of UAVs. To address this problem, adversarial training and defensive distillation methods are further investigated and evaluated, increasing the robustness of deep learning models in UAVs. To our knowledge, this is the first study on adversarial attacks and defenses against deep learning-based UAVs, which calls for more attention to the security and safety of such safety-critical applications.

Deep neural networks have been applied to many tasks in air traffic management, ranging from anomaly detection to flight trajectory prediction. However, it has been shown that such algorithms are susceptible to adversarial examples. In this letter, we manage to show that current deep learning algorithms proposed for spoofing detection are vulnerable to maliciously crafted ADS-B data. To inject the false messages into the ADS-B system without being detected, we need to find adversarial perturbations to balance the need of overwhelming the channel noise and keeping the decoding error low. Simulation results demonstrate the viability of our approach to evade the DNN-based spoofing detector without increasing the decoding error.

With the fast increase in airspace density and high-safety requirements on aviation, automatic dependent surveillance-broadcast (ADS-B) is regarded as the primary method in the next generation air traffic surveillance. The ADS-B data is broadcast with the plain text without sufficient security measures, which results in various attack patterns emerging. However, in terms of constrictions with laws and regulations, ADS-B attack data is difficult to collect and obtain, which is essential for data security research studies. To deal with the absence of ADS-B attack data in real environments, the construction strategy on ADS-B attack data is proposed. For construction fidelity, ADS-B data features are analysed and modelled at first. Then the popular and classical attack patterns on ADS-B data are analysed to establish threat models. Based on the original ADS-B data sets, the construction strategy is designed to focus on attack target selection, key parameter determination, and mixture strategy, reproducing the attack intentions. The constructed ADS-B attack data sets are hybrid data sets including the normal and attack data. By simulation analyses, the feasibility and availability of the construction strategy were validated with real ADS-B data.

Coverage-guided grey-box fuzzing for computer systems has been explored for decades. However, existing techniques do not adequately explore the space of continuous behaviors in Cyber-Physical Systems (CPSs), which may miss safety-critical bugs. Optimization-guided falsification is promising to find violations of safety specifications, but not suitable for identifying traditional program bugs. This article presents a fuzzing process for finding safety violations at the development phase, which is guided by two quantities: a branch coverage metric to explore discrete program behaviors and a Linear Temporal Logic (LTL) robust satisfaction metric to identify undesirable continuous plant behaviors. We implement CPFuzz to demonstrate the utility of the idea and estimate its effectiveness on seven control system benchmarks. The results show up to a better performance in average time to find violations on all benchmarks than S-TaLiRo and six benchmarks than S3CAMX. Finally, we exploit CPFuzz to synthesize the sensor spoofing attack on a DC motor with fixed-point overflow vulnerability as a case study.

Vulnerability of various machine learning methods to adversarial examples has been recently explored in the literature. Power systems which use these vulnerable methods face a huge threat against adversarial examples. To this end, we first propose a more accurate and computationally efficient method called Adaptive Normalized Attack (ANA) to attack power systems using generate adversarial examples. We then adopt adversarial training to defend against attacks of adversarial examples. Experimental analyses demonstrate that our attack method provides less perturbation compared to the state-of-the-art FGSM (Fast Gradient Sign Method) and DeepFool, while our proposed method increases misclassification rate of learning methods for attacking power systems. In addition, the results show that the proposed adversarial training improves robustness of power systems to adversarial examples compared to using state-of-the-art methods.

With the requirements on accuracy, coverage and reliability, the air traffic surveillance is being developed into the next generation. In 2020, ADS-B data is becoming the foundation to establish air traffic situation awareness capabilities. However, ADS-B is designed without sufficient security guarantees, which results in diverse attack threats. Hence, it is in demand of effective attack detections to keep attack data away from decision flows. To improve the accuracy and robustness, attack detection based on generative adversarial network for ADS-B data is proposed. The LSTM networks are the core components to set up the generator and discriminator to make the most of temporal spatial correlations. Utilizing the reconstruction error and discriminative loss, the comprehensive detection metric is obtained to identify attack behaviors. To enhance the robustness, the analysis threshold for detection decision is determined in terms of normal data intrinsic features. By experimental analysis on real ADS-B data, the accuracy and robustness of the proposed method is validated.

Power grid cyber-physical systems face a variety of cyber-attacks. Machine learning based stealthy false data injection attacks are explored in this paper. To avoid the detection of the Bad Data Detector, two issues are considered in the machine learning based attack strategy. One is how to use machine learning to generate suitable attack vectors based on eavesdropping measurements. The other is how to improve the robustness of the attack strategy if there are corrupted measurements or outliers. Considering these two problems, a robust linear regression based false data injection attack strategy is proposed. Moreover, a more sparse attack strategy is also explored to further reduce the cost of attackers. Simulations conducted on the IEEE 14-bus system verify the effectiveness of the attack strategies.

For the next generation air traffic surveillance, ADS-B is becoming the primary method to obtain more accurate data with wide coverage, which establishes the foundation for automatic and intelligent air traffic management system. However, ADS-B is designed without sufficient security considerations, transmitting with plain text without integrity and authentication validations. Thus, ADS-B data is in face of various attack threats, which may cause disruptions on system availability and reliability. To eliminate effects of attack behaviours, attack detection is in demand to avoid attack data injecting into decision making flow. Based on hidden Markov model with sticky hierarchical Dirichlet process, the dynamic temporal detection method is proposed to detect multiple attack patterns. Taking advantage of multiple attribute data, the dimensions of data are reduced to one dimension to set up feature sequences. With sticky hierarchical Dirichlet process, the parameters are obtained for hidden Markov model dynamically. Utilizing hidden Markov model, the generative model is established to predict hidden states of ADS-B data sequence. By analysing the contextual deviation information on hidden state sequences, the attack behaviours are discriminated to determine the attack data. By experiments on real ADS-B data, the feasibility and accuracy of the proposed method are validated.

In the next generation air traffic surveillance, ADS-B is the primary surveillance method to improve situation awareness capabilities. But ADS-B protocol is absent of sufficient security considerations, especially for data integrity and authentication. As a result, attack patterns on ADS-B data are emerging and efficient attack detection strategies are in great demand to enhance data security. To decrease the time delay of detection, enhance accuracy and mitigate the concept drift impacts, the online sequential attack detection strategy based on hierarchical temporal memory are proposed. By applying binary encoding, ADS-B data is transformed into sparse distribution representation with temporal and spatial correlations. The encoded data is push into hierarchical temporal memory and online learning schemes are established for the ADS-B stream data. With the sequential ADS-B data, hierarchical temporal memory is updated and used to generate the deviations between predictions and original values for the corresponding ADS-B data. Designing and applying deviation analysis, sequential analysis and adaptive threshold check, the differences between normal and novelty are magnified and easy to be distinguished. According to experimental analysis, the attack detection strategy is efficient on processing time and accuracy.