Search for: All records

Deep learning-based semantic communication systems leverage neural networks to efficiently extract and transmit semantic information. However, current semantic communication systems still face potential physical-layer security problems, including backdoor attacks with physical-layer triggers and adversarial attack signals. In this work, we first identify a critical physical-layer security vulnerability in semantic communication systems. Specifically, an attacker can covertly implant a backdoor into a model without requiring direct access to the training pipeline, dataset, or model parameters. Consequently, the receiver’s inference outputs can be maliciously manipulated via the attacker’s trigger signal. To realize this attack, an over-the-air backdoor injection (OABI) algorithm is proposed. In the training phase, the attacker injects an iteratively updated trigger into the channel to implant the backdoor. To defend against backdoor injection during training, a proactive defense strategy is proposed by constructing a subspace on the received signal to capture components that persistently decrease the loss. The simulation results demonstrate the effectiveness of the proposed attack, and the attack success rate is improved by up to 55.44% over the baseline. In addition, the proposed defense method improves model performance and mitigates backdoor injection threats. 

Deep learning-based semantic communication systems leverage neural networks to efficiently extract and transmit semantic information. However, these systems still face physical-layer security risks that have not been fully addressed. In this work, we first identify a critical vulnerability in semantic communication systems that allows an attacker to covertly implant a backdoor into the receiver model over the air without direct access to the training pipeline, dataset, or model parameters. Once the backdoor is injected, the receiver’s inference outputs can be maliciously manipulated by transmitting an attacker-designed trigger signal. To study this threat, an over-the-air backdoor injection (OABI) algorithm is proposed. During training, the attacker injects an iteratively updated trigger into the wireless channel to implant the backdoor. To mitigate this threat, a proactive defense strategy is further proposed, in which suspicious signals are identified by a lightweight probe neural network and attack-related components are suppressed during training. Simulation results demonstrate the effectiveness of the proposed attack, with the attack success rate improving by up to 41.98% over the baseline. The results also show that existing defense baselines fail to prevent OABI, whereas the proposed defense effectively mitigates the backdoor injection threat and preserves normal model performance. 

A vehicular communication network allows vehicles on the road to be connected by wireless links, providing road safety in vehicular environments. Vehicular communication network is vulnerable to various types of attacks. Cryptographic techniques are used to prevent attacks such as message modification or vehicle impersonation. However, cryptographic techniques are not enough to protect against insider attacks where an attacking vehicle has already been authenticated in the network. Vehicular network safety services rely on periodic broadcasts of basic safety messages (BSMs) from vehicles in the network that contain important information about the vehicles such as position, speed, received signal strength (RSSI) etc. Malicious vehicles can inject false position information in a BSM to commit a position falsification attack which is one of the most dangerous insider attacks in vehicular networks. Position falsification attacks can lead to traffic jams or accidents given false position information from vehicles in the network. A misbehavior detection system (MDS) is an efficient way to detect such attacks and mitigate their impact. Existing MDSs require a large amount of features which increases the computational complexity to detect these attacks. In this paper, we propose a novel grid-based misbehavior detection system which utilizes the position information from the BSMs. Our model is tested on a publicly available dataset and is applied using five classification algorithms based on supervised learning. Our model performs multi-classification and is found to be superior compared to other existing methods that deal with position falsification attacks.