Machine unlearning aims to eliminate the influence of specific data from trained models to ensure privacy compliance. However, most existing methods assume full access to the original training dataset, which is often impractical. We address a more realistic yet challenging setting: few-shot zero-glance, where only a small subset of the retained data is available and the forget set is entirely inaccessible. We introduce GFOES, a novel framework comprising a Generative Feedback Network (GFN) and a two-phase fine-tuning procedure. GFN synthesises Optimal Erasure Samples (OES), which induce high loss on target classes, enabling the model to forget class-specific knowledge without access to the original forget data, while preserving performance on retained classes. The two-phase fine-tuning procedure enables aggressive forgetting in the first phase, followed by utility restoration in the second. Experiments on three image classification datasets demonstrate that GFOES achieves effective forgetting at both logit and representation levels, while maintaining strong performance using only 5% of the original data. Our framework offers a practical and scalable solution for privacy-preserving machine learning under data-constrained conditions.

Facing the complexity of escape routing in high-density interconnects within chiplet systems, conventional methods often struggle to balance efficiency and quality under large-scale design scenarios. This paper presents a Sliding-Window-based progressive Integer linear Programming (ILP) for scalable Escape Routing of chiplet interconnects, called SWIPER, which maintains high routing quality with controllable complexity by sequentially optimizing localized subproblems. Our approach uses predefined geometric patterns to adaptively select single- or multi-layer paths using ILP model, to avoid routing conflicts, thereby improving routing flexibility. Experimental results demonstrate that, compared with existing methods, SWIPER improves the routing completion rate by up to 28% in obstacle-dense scenarios, reduces the number of vias by up to 18%, and reduces the wirelength by up to 1.5%, exhibiting excellent scalability and practical robustness.

Link flooding attacks (LFA) represent a destructive yet stealthy type of Distributed Denial of Service (DDoS) attack that overwhelms target inter-satellite links (ISLs) within low Earth orbit (LEO) satellite networks. However, existing LFA detection and mitigation solutions, while effective in terrestrial networks, are not well-suited to address LFAs in satellite networks due to their unique characteristics. To address this challenge, we propose SatLFAGuard, a novel framework for LFA detection and mitigation. The core innovation lies in reformulating the LFA detection problem as a classification task to identify adjacent abnormal nodes. This classification problem can be effectively tackled using spatial and temporal graph convolutional neural network, which leverages features that capture both satellite node-level information and the spatial topology of LEO networks. For LFA mitigation, we introduce a multipath coordinated transmission strategy that schedules multiple paths within LEO and transmit the congested traffic in parallel, compelling attackers to abandon the attack by substantially increasing their resource cost. Specifically, SatLFAGuard comprises three key components: (1) a satellite node state awareness mechanism based on In-band Network Telemetry (INT), which periodically collects real-time, fine-grained node load levels; (2) a graph-based anomaly detection module leveraging an attention-enhanced spatiotemporal graph convolutional network (ASTGCN) to identify abnormal nodes and (3) a multipath coordinated transmission strategy that dispatches traffic across multiple paths upon attack detection, increasing the cost for adversaries while preserving legitimate communication. The experimental results on the Iridium constellation demonstrate that SatLFAGuard can accurately detect LFA attacks while maintaining a low false positive rate, significantly outperforming than baseline methods. Upon the detection of LFA, SatLFAGuard is able to assimilate the attack traffic in a responsive manner with acceptable resource overhead.

As voice assistants (VAs) become increasingly popular, concerns about their privacy and security have garnered significant attention. VAs nowadays rely on voiceprint authentication to enhance their security. However, this method is susceptible to spoofing attacks, where attackers may use recording or synthesis techniques to mimic the user's voice, thereby bypassing the authentication mechanism. To address this, we introduce “BoneAuth,” a novel liveness detection system in this article. It offers continuous voice authentication for users, enhancing the security of VAs. BoneAuth is designed to be used in wearable devices with built-in microphones, such as Bluetooth earphones. Our basic idea is continuously matching the user's voice signals with the vibration signals produced by their vocal cords during speech. Specifically, our system uses the device's built-in microphone to concurrently capture vibrations from bone conduction (BC) and voices from air conduction (AC). We introduce a signal separation algorithm that, by measuring the unique threshold range of the user, can separate the AC and BC signals from the mixed microphone signals. By continuously comparing the consistency of the two signals, our system can determine whether the user's voice is a real live voice or artificially generated voice. Our system does not require user-specific passphrases for authentication, making it easy to deploy and use without the need for additional user actions or hardware. We demonstrate the feasibility of our method using commercial off-the-shelf Bluetooth earphones. Extensive experiments show an accuracy rate close to 98.15%, proving the effectiveness of our approach.

The right to deletion mandates that data controllers receiving deletion requests not only erase the specified data but also notify other controllers to do the same. Designing a secure and privacy-preserving protocol for distributing deletion notifications across involved data controllers, a topic not previously addressed in the literature, presents significant challenges. In this paper, we introduce CAREFUL, a secure and privacy-preserving deletion notification distribution protocol—the first to address these challenges. It operates within a centralized architecture consisting of regulatory and service planes. The fundamental principle of CAREFUL is that the regulatory plane creates a cryptographic access control structure, ensuring that data controllers in the service plane only identify the next-hop nodes for notification. With CAREFUL, the circulation history of data pending for deletion, which is a special user’s privacy, can be preserved. Moreover, CAREFUL protects the deletion notification distribution from various malicious attacks over untrusted underlay networks. Experimental results validate CAREFUL’s practicality and its efficiency in terms of resource overhead.

Link Flooding Attacks (LFA) are increasingly challenging the availability and stability of Software-Defined Networks (SDN), leveraging their distributed and covert nature to escape detection. Existing methods such as packet analysis and behavioral patterns struggle with pinpointing specific target attack paths due to the dynamic routing and distribution of attack traffic across multiple entry points, making it difficult to implement defensive measures effectively. This paper introduces a novel detection method using a Spatial-Temporal Graph Convolutional Network (STGCN), which combines the strengths of Convolutional Neural Networks (CNN) for temporal pattern recognition and Graph Neural Networks (GNN) for spatial topology analysis. Unlike traditional approaches, our model leverages In-Band Network Telemetry (INT) for real-time traffic and topology monitoring, enhancing our ability to pinpoint and mitigate LFAs. We innovatively transform the detection of attacked links into a dual-point anomaly detection problem, significantly increasing the accuracy of identifying compromised links. Experimental results demonstrate that our method not only achieves high detection accuracy but also ensures the efficient use of network resources, making it particularly effective for resource-constrained environments.

Continuous authentication is crucial for protecting user’s privacy throughout their login session. Existing studies employ wireless sensing technologies to provide device-free and unobtrusive authentication; the user’s behavior is continually assessed without their direct involvement until it deviates from their normal pattern. However, these works primarily concentrate on single-user authentication, which poses challenges in multiuser scenarios, such as smart homes and offices, where more than one user usually exists. In this article, we propose HeartPrint, a continuous multiuser authentication system, that employs a single commodity mmWave radar to capture the unique self-driving heartbeat motions from multiple users. Specifically, HeartPrint leverages the effect of skin surface vibrations caused by heartbeat on radio frequency (RF) transmissions. To profile individual heartbeat signals from the entangled components that are induced by multiple users, we first use a clustering method to position each user in the environment, then focus on the signal reflected from each position separately. The irrelevant body movements are eliminated from the RF signal by using a proposed signal energy comparison method for preserving fine-grained heartbeat traits. We then develop a pipeline to extract the most informative features for characterizing each user and feed them to an elaborated classifier for user authentication. We evaluate HeartPrint with 54 participants and demonstrate that it achieves an average authentication accuracy of over 95%. Additionally, we show that it is resilient against spoofing attacks, with an average attack success rate of less than 3%.

Domain generation algorithms (DGAs) are widely used in modern botnets to generate a large number of domain names through which bots can communicate with their command and control (C & C) servers. In recent years, many machine learning based approaches have been proposed to automatically detect algorithmically generated domains in real time and have achieved success in traditional DGAs. Nevertheless, they are somewhat unavailable for adversarial domains. In this paper, we develop a more threatening DGA called ShadowDGA that utilizes generative adversarial networks (GANs) to simulate the distribution of benign domains without any knowledge about the DGA detector to evade detection. Experimental results demonstrate that the domains generated by ShadowDGA are the most difficult to detect compared to existing DGA families. We also present an effective defense method for adversarial domains without retraining. These findings indicate that detectors that rely solely on features extracted from the domain name are vulnerable, while a robust DGA detector should contain additional contextual information.

The \emph{Locator/Identifier Separation Protocol} (LISP) introduces several benefits to the Internet architecture, yet, since it is just in the initial deployment stage, comprehensive understanding of its integration performance with legacy Internet becomes essential. We leverage on RIPE Atlas, the largest Internet measurement infrastructure, to conduct large scale measurements analysis to provide the feedback to improve LISP technology. The preliminary evaluations show that LISP generally has a reliable performance, compared with the existing Internet. From our vantage point, we observe that LISP introduces a non-negligible latency for the European and North American destinations, occasionally some extremely large delay, however, it shows a faster connection for the Asian intercontinental transmission.