Smart Internet of Things (IoT) devices generate vast, distributed data, and their limited computational and storage capacities complicate data protection. Federated Learning (FL) enables collaborative model training across clients, enhancing performance and protecting data privacy. The Right to be Forgotten (RTBF) raises the demand for precise data removal. Federated Unlearning (FU) offers a solution for accurate data deletion in FL systems. Existing FU methods often struggle to simultaneously ensure effective data forgetting and preserve model generalization. To mitigate these challenges, an effective and robust FU framework has been proposed, which is based on the “Bad Teacher” knowledge distillation (KD), termed FedBT. First, the “Bad Teacher" KD guides the trained model to eliminate specific client contributions from the global model. Next, the frequency domain extracts the global model’s generalization components. Finally, orthogonal constraints are applied to the KD-generated gradients within the orthogonal subspace of these components, ensuring the gradients preserve the trained model’s generalization ability. FedBT eliminates the need to store historical records of parameter updates. Using orthogonal space constraints, the generalization ability of the trained model is safeguarded during unlearning. Extensive experiments on three datasets with various metrics show our method reduces accuracy by only 0.53% on MNIST, 0.26% on Fashion-MNIST, and 4.67% on CIFAR10, surpassing the best approach. Furthermore, FedBT obtains an unlearning performance that most closely approximates the results obtained from retraining from scratch. FedBT boosts IoT security by enabling the “forgetting" of certain client data, crucial for protecting user privacy and ensuring secure device interactions.

Speaker Recognition Systems (SRSs) are becoming increasingly popular in various aspects of life due to advances in technology. However, these systems are vulnerable to cyber threats, particularly adversarial attacks. Traditional adversarial attack methods, such as the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD), are designed for a white-box setting where attackers have complete knowledge of the inner workings of the target systems. This limits the practicality of these attacks. To overcome this limitation, we propose a new attack model that uses a neural network to generate adversarial examples directly, without the need for full knowledge of the recognition model in a target SRS. In addition, we have designed a novel loss function to balance the effectiveness and confidentiality of adversarial examples. Our new approach was evaluated against SincNet, a state-of- the-art SRS. Experimental results show that our approach achieves outstanding performance, with the best attack success rate of 99.83% and the best Signal-to-Noise Ratio (SNR) value of 41.30.

Internet of Things (IoT) is fast growing. Non-personal computer devices under the umbrella of IoT have been increasingly applied in various fields and will soon account for a significant share of total Internet traffic. However, the security and privacy of IoT and its devices have been challenged by malware, particularly polymorphic worms that rapidly self-propagate once being launched and vary their appearance over each infection to escape from the detection of signature-based intrusion detection systems. It is well recognized that polymorphic worms are one of the most intrusive threats to IoT security. To build an effective, strong defense for IoT networks against polymorphic worms, this study proposes a machine intelligent system, termed Gram-Restricted Boltzmann Machine (Gram-RBM), which automatically generates generic fingerprints/signatures for the polymorphic worm. Two augmented N-gram-based methods are designed and applied in the derivation of polymorphic worm sequences, also known as fingerprints/signatures. These derived sequences are then optimized using the Gaussian–Bernoulli RBM dimension-reduction algorithm. The results, gained from the experiments involved three different types of polymorphic worms, show that the system generates accurate fingerprints/signatures even under “noisy” conditions and outperforms related methods in terms of accuracy and efficiency.

With the rapid increase in the amount and type of malware, traditional methods of malware detection and family classification for IoT applications through static and dynamic analysis have been greatly challenged. In this paper, a new simple and effective attention module of Convolutional Neural Networks (CNNs), named as Depthwise Efficient Attention Module (DEAM), is proposed and combined with a DenseNet to propose a new malware detection and family classification model. Based on the good effect of the DenseNet in the field of image classification and the visual similarity of the malware family on images, the gray-scale image transformed from malware is input into the model combined with the DEAM and DenseNet for malware detection, and then the family classification is carried out. The DEAM is a general lightweight attention module improved based on the Convolutional Block Attention Module (CBAM), which can strengthen the attention to the characteristics of malware and improve the model effect. We use the MalImg dataset, Microsoft malware classification challenge dataset (BIG 2015), and our dataset constructed by the two above-mentioned datasets to verify the effectiveness of the proposed model in family classification and malware detection. Experimental results show that the proposed model achieves 99.3% in terms of accuracy for malware detection on our dataset and achieves 98.5% and 97.3% in terms of accuracy for family classification on the MalImg dataset and BIG 2015 dataset, respectively. The model can reliably detect IoT malware and classify its families.

5G is about to open Pandora’s box of security threats to the Internet of Things (IoT). Key technologies, such as network function virtualization and edge computing introduced by the 5G network, bring new security threats and risks to the Internet infrastructure. Therefore, higher detection and defense against malware are required. Nowadays, deep learning (DL) is widely used in malware detection. Recently, research has demonstrated that adversarial attacks have posed a hazard to DL‐based models. The key issue of enhancing the antiattack performance of malware detection systems that are used to detect adversarial attacks is to generate effective adversarial samples. However, numerous existing methods to generate adversarial samples are manual feature extraction or using white‐box models, which makes it not applicable in the actual scenarios. This paper presents an effective binary manipulation‐based attack framework, which generates adversarial samples with an evolutionary learning algorithm. The framework chooses some appropriate action sequences to modify malicious samples. Thus, the modified malware can successfully circumvent the detection system. The evolutionary algorithm can adaptively simplify the modification actions and make the adversarial sample more targeted. Our approach can efficiently generate adversarial samples without human intervention. The generated adversarial samples can effectively combat DL‐based malware detection models while preserving the consistency of the executable and malicious behavior of the original malware samples. We apply the generated adversarial samples to attack the detection engines of VirusTotal. Experimental results illustrate that the adversarial samples generated by our method reach an evasion success rate of 47.8%, which outperforms other attack methods. By adding adversarial samples in the training process, the MalConv network is retrained. We show that the detection accuracy is improved by 10.3%.

Abstract Internet worms can self‐propagate over the Internet, and have caused significant damages to the Internet infrastructure. To speed up the propagating process, the worms need to scan many Internet Protocol (IP) addresses to target vulnerable hosts. However, the distribution of IP addresses is highly non‐uniform, which results in many scans wasted on invulnerable addresses. Inspired by the theory of good point set, this paper proposes a new scanning strategy, referred to as good point set scanning (GPSS), for worms. Experimental results show that GPSS can generate more distinct IP addresses and less unused IP addresses than the permutation scanning. Combined with group distribution, a static optimal GPSS is derived. Since the information cannot be easily collected before a worm is released, a self‐learning worm with GPSS is designed. Such worm can accurately estimate the underlying vulnerable‐host distribution when a sufficient number of IP addresses of infected hosts are collected. We use a modified Analytical Active Worm Propagation (AAWP) to simulate data of Code Red and the performance of different scanning strategies. Experimental results show that once the distribution of vulnerable hosts is accurately estimated, a self‐learning worm can propagate much faster than other worms. Finally, some possible countermeasures are given. Copyright © 2008 John Wiley & Sons, Ltd.