Phishing attacks are a major threat to online security, exploiting user vulnerabilities to steal sensitive information. Various methods have been developed to counteract phishing, each with varying levels of accuracy, but they also face notable limitations. In this study, we introduce PhishAgent, a multimodal agent that combines a wide range of tools, integrating both online and offline knowledge bases with Multimodal Large Language Models (MLLMs). This combination leads to broader brand coverage, which enhances brand recognition and recall. Furthermore, we propose a multimodal information retrieval framework designed to extract the relevant top k items from offline knowledge bases, using available information from a webpage, including logos and HTML. Our empirical results, based on three real-world datasets, demonstrate that the proposed framework significantly enhances detection accuracy and reduces both false positives and false negatives, while maintaining model efficiency. Additionally, PhishAgent shows strong resilience against various types of adversarial attacks.

Current unlearning methods for large language models usually rely on reverse optimization to reduce target token probabilities. However, this paradigm disrupts the subsequent tokens prediction, degrading model performance and linguistic coherence. Moreover, existing evaluation metrics overemphasize contextual forgetting while inadequately assessing response fluency and relevance. To address these challenges, we propose ReLearn, a data augmentation and fine-tuning pipeline for effective unlearning, along with a comprehensive evaluation framework. This framework introduces Knowledge Forgetting Ratio (KFR) and Knowledge Retention Ratio (KRR) to measure knowledge-level preservation, and Linguistic Score (LS) to evaluate generation quality. Our experiments show that ReLearn successfully achieves targeted forgetting while preserving high-quality outputs. Through mechanistic analysis, we further demonstrate how reverse optimization disrupts coherent text generation, while ReLearn preserves this essential capability.

Recent progress in Multimodal Large Language Models (MLLMs) has unlocked powerful cross-modal reasoning abilities, but also raised new safety concerns, particularly when faced with adversarial multimodal inputs. To improve the safety of MLLMs during inference, we introduce a modular and adaptive inference-time intervention technology, AutoSteer, without requiring any fine-tuning of the underlying model. AutoSteer incorporates three core components: (1) a novel Safety Awareness Score (SAS) that automatically identifies the most safety-relevant distinctions among the model’s internal layers; (2) an adaptive safety prober trained to estimate the likelihood of toxic outputs from intermediate representations; and (3) a lightweight Refusal Head that selectively intervenes to modulate generation when safety risks are detected. Experiments on LLaVA-OV and Chameleon across diverse safety-critical benchmarks demonstrate that AutoSteer significantly reduces the Attack Success Rate (ASR) for textual, visual, and cross-modal threats, while maintaining general abilities. These findings position AutoSteer as a practical, interpretable, and effective framework for safer deployment of multimodal AI systems.

Traditional Network Intrusion Detection Systems (NIDS) often generate large volumes of alerts with redundancies and false positives, incapable of correlating detected attack actions. This adds difficulty for security analysts to construct a comprehensive understanding of multi-step attacks. To address these limitations, we present a novel MITRE-based Multi-step Attack Scenario Construction (M2ASK) algorithm that enhances cyber threat intelligence (CTI) by integrating MITRE ATT&CK tactic and technique mapping, facilitating the interpretation of multi-step attacks and informing response strategies. Our approach processes alert data from NIDSs, transforming it into a network communication graph. Graph-based correlation techniques are employed, combined with MITRE ATT&CK and Cyber Kill Chain stage profiling to construct comprehensive network attack scenarios. Our key contributions include: (1) the development of a Cyber Kill Chain based model for constructing attack scenarios; (2) the alert correlation approach based on MITRE ATT&CK tagging of attack actions.

As Information Technology (IT) infrastructures have become increasingly complex to secure against accelerating cyber threats, current threat detection approaches have been largely silos in nature; security analysts in the environment are typically bombarded with large volume of security alerts that often cause severe fatigues and the possibility of judgement errors. This problem is further exacerbated by the number of false-positives that analysts may waste valuable time and resources pursuing. In this paper, we present how intuitive graph-based machine learning can be used to address the problem of alert fatigue and prioritize risky alerts to assist security analysts. The rationale and workflow of the proposed Graph Analysis (GA) algorithm is discussed in detail, with its effectiveness demonstrated by simulated experiments.

Nowadays, portable devices demand small-sized and low-end loudspeaker, however, the physical acoustic bass (or low-frequency) reproduction is usually poor. Bass enhancement by equalization is not feasible, and may even overload or damage the loudspeaker. Bass enhancement for such low-end loudspeaker can be psychoacoustically accomplished by exploiting the missing fundamental phenomenon. These systems are known as virtual bass systems (VBS), and generally use nonlinear device (NLD) as the main processing block to generate harmonics for virtual pitch perception. One of the recently developed NLD is the Arc-Tangent Square Root (ATSR) function, which can be used to control the pitch perception by a set of parameters. Mathematical relation between the input and output of the NLD is derived under a single-tone analysis framework. A detailed study on how the parameters of this NLD affect the harmonic's decay pattern and intensity is presented in this paper.

Low frequency bandwidth limitation is a common problem faced by miniature speakers and highly-directional speakers that have relatively high cut-off frequency. As such, low frequencies cannot be effectively reproduced. One of the current methods in addressing this problem is to employ psychoacoustic signal processing based on the “missing fundamental phenomenon”. Nonlinear device is generally used to induce virtual pitch that enhances the low frequency perception. However, some of the difficulties in using nonlinear device include the need of precise adjustment of harmonics' magnitudes, harmonic order and its decay rate to achieve good perceived bass. In this paper, we propose two techniques on synthesis of polynomial-based nonlinear device and harmonic shifting by modulation in an attempt to overcome these difficulties. Real-Time implementation and listening tests were conducted to verify the effectiveness of the proposed algorithm.