Face recognition systems have become ubiquitous in biometric authentication and surveillance, yet our understanding of how deep neural networks encode facial identity remains incomplete. While human face perception integrates multiple cues (geometric shape, surface texture, and local features), the relative importance of these signals in machine learning models is poorly characterized. This work addresses a fundamental question: do modern deep face recognition models rely primarily on geometric shape or surface texture for identity matching? To answer this, we conducted a controlled empirical study using synthetic face pairs with systematically manipulated shape and texture information. Our results reveal a striking and consistent pattern: all evaluated models exhibit texture bias. Spatial attribution analysis using occlusion-based heatmaps reveals that texture-based identity relies heavily on periocular regions, while shape-based identity is more broadly distributed. These findings have significant implications for understanding model robustness, fairness, and the interpretability of deployed face recognition systems. Our methodology provides a reproducible framework for quantifying and visualizing how deep neural networks decompose facial identity signals, contributing to the growing effort to build trustworthy biometric systems.

Deepfake speech detectors often output a single score without explaining why an audio sample is flagged, where in the signal the evidence lies, or what cues drive the decision. We propose an audio-native explainability pipeline using Integrated Gradients on time-aligned self-supervised representations to localize decision evidence over time. We apply the proposed method to three WavLM-based detectors (AASIST, CA-MHFA, SLS) on ASVspoof 5 and manually annotate the highest-attribution regions to provide a semantic meaning of the most important cues. Despite similar performance, the detectors rely on different cues: AASIST emphasizes non-speech/environment cues, CA-MHFA focuses on localized phoneme artifacts, and SLS relies on word boundaries and spectral integrity. We move beyond speculative reasoning and validate our findings by causal masking of the primary detector cues. Observed performance degradation further supports the explained detector semantics.

Large language models (LLMs) are vulnerable to jailbreak attacks that bypass safety alignment through carefully crafted prompts. Many existing defenses require access to model weights or internals, making them difficult to apply to black-box deployments. We propose AlcaTRAz (Anchored Tree-Rule defense Against jailbreaks), a prompt-level defense based on rule trees that operates exclusively on the input text and requires no modification or retraining of the target model. The method automatically learns a transferable transformation rule that inserts controlled character-level perturbations at selected positions, thereby disrupting structural regularities exploited by jailbreak attacks while largely preserving the model's utility on benign queries. We evaluate the proposed method across 33 open-weight models, 22 jailbreak attack types, and a benchmark of short, single-turn benign questions, comparing against three representative prompt-level baselines (Llama Guard, RA-LLM, Goal Prioritization). Among the compared defenses, AlcaTRAz achieves the best composite security and functionality score in 73.4

We introduce a spoofing countermeasure architecture conditioned on speaker-reference recordings, but observe that it converges to a solution that effectively ignores the reference during inference. Surprisingly, training with a reference channel induces invariance that improves deepfake detection, even when the reference is absent or mismatched during inference. Based on this observation, we propose a Reference-Augmented Training (RAT) strategy. RAT yields improved detection performance compared to single-utterance baselines, even when the reference recording is replaced with a zero vector at inference. Through rigorous analysis, we demonstrate that the optimization process rapidly diminishes the reference contributions, leading to inference largely independent of the reference channel. Using RAT, we achieve state-of-the-art 2.57

Synthetic speech detection benchmarks now report sub-1

While deepfake speech detectors built on large self-supervised learning (SSL) models achieve high accuracy, employing standard ensemble fusion to further enhance robustness often results in oversized systems with diminishing returns. To address this, we propose an evolutionary multi-objective score fusion framework that jointly minimizes detection error and system complexity. We explore two encodings optimized by NSGA-II: binary-coded detector selection for score averaging and a real-valued scheme that optimizes detector weights for a weighted sum. Experiments on the ASVspoof 5 dataset with 36 SSL-based detectors show that the obtained Pareto fronts outperform simple averaging and logistic regression baselines. The real-valued variant achieves 2.37

Claims about the robustness and fairness of deepfake speech detectors are only as credible as the datasets used to train and evaluate those systems. We present a dataset-level audit of the deepfake speech landscape. We compile and analyze 39 deepfake speech datasets, examining key attributes including accessibility, documentation, demographic and language coverage, dataset scale, and the underlying bona fide speech sources. Our audit reveals two important takeaways. Firstly, fairness assessment is largely infeasible because most datasets lack demographic metadata, and only a few contain gender or language labels. This prevents any meaningful subgroup analysis and leaves other demographic attributes unaddressed. Secondly, we identify substantial overlap in underlying bona fide source corpora across datasets, which can undermine cross-dataset evaluation and lead to overstated generalization claims.

Open-set source tracing is increasingly framed as a verification problem, motivating the use of pairwise metric-learning objectives from biometrics. We thus compare global anchoring and pairwise verification under matched backbones and a fixed data and epoch budget on MLAAD (in-domain) and STOPA (out-of-domain). In our runs, global anchoring yields lower in-domain error (8.61

Advances in speech synthesis have made deepfake audio highly realistic. Earlier studies reported 70-80

We present SpAArSIST, a deployment-oriented refinement of the widely used AASIST graph pooling backend for self-supervised learning (SSL) based anti-spoofing. Motivated by redundant operations in public implementations, we replace learned pooling and stack-node attention with explicit, lightweight choices: separate train and inference graph pooling ratios (k_tr,k_inf), magnitude-based node scoring, and mean aggregation of graph nodes. The best overall configuration (rank 1) cuts backend compute by 20.7

Despite growing attention to deepfake speech detection, the aspects of bias and fairness remain underexplored in the speech domain. To address this gap, we introduce the Speaker Characteristics Deepfake (SCDF) dataset: a novel, richly annotated resource enabling systematic evaluation of demographic biases in deepfake speech detection. SCDF contains over 237,000 utterances in a balanced representation of both male and female speakers spanning five languages and a wide age range. We evaluate several state-of-the-art detectors and show that speaker characteristics significantly influence detection performance, revealing disparities across sex, language, age, and synthesizer type. These findings highlight the need for bias-aware development and provide a foundation for building non-discriminatory deepfake detection systems aligned with ethical and regulatory standards.

Modern search engines extensively personalize results by building detailed user profiles based on query history and behaviour. While personalization can enhance relevance, it introduces privacy risks and can lead to filter bubbles. This paper proposes and evaluates a lightweight, client-side query obfuscation strategy using randomly generated multilingual search queries to disrupt user profiling. Through controlled experiments on the Seznam.cz search engine, we assess the impact of interleaving real queries with obfuscating noise in various language configurations and ratios. Our findings show that while displayed search results remain largely stable, the search engine's identified user interests shift significantly under obfuscation. We further demonstrate that such random queries can prevent accurate profiling and overwrite established user profiles. This study provides practical evidence for query obfuscation as a viable privacy-preserving mechanism and introduces a tool that enables users to autonomously protect their search behaviour without modifying existing infrastructure.

Facial occlusions in surveillance footage can obscure important features, preventing facial recognition systems from identifying people. This work focuses on reconstructing these missing facial parts using Generative Adversarial Networks (GANs) to improve facial recognition accuracy while maintaining a low false acceptance rate. Additionally, we investigate how the generated images can be further enhanced using various image enhancement methods to boost recognition accuracy. To evaluate the results, we conduct experiments with widely used face embedding models, such as QMagFace and ArcFace, to determine whether image reconstruction and enhancement improve face recognition accuracy.

A key research area in deepfake speech detection is source tracing - determining the origin of synthesised utterances. The approaches may involve identifying the acoustic model (AM), vocoder model (VM), or other generation-specific parameters. However, progress is limited by the lack of a dedicated, systematically curated dataset. To address this, we introduce STOPA, a systematically varied and metadata-rich dataset for deepfake speech source tracing, covering 8 AMs, 6 VMs, and diverse parameter settings across 700k samples from 13 distinct synthesisers. Unlike existing datasets, which often feature limited variation or sparse metadata, STOPA provides a systematically controlled framework covering a broader range of generative factors, such as the choice of the vocoder model, acoustic model, or pretrained weights, ensuring higher attribution reliability. This control improves attribution accuracy, aiding forensic analysis, deepfake detection, and generative model transparency.

Proof-of-Stake (PoS) consensus protocols often face a trade-off between performance and security. Protocols that pre-elect leaders for subsequent rounds are vulnerable to Denial-of-Service (DoS) attacks, which can disrupt the network and compromise liveness. In this work, we present PoS-CoPOR, a single-chain PoS consensus protocol that mitigates this vulnerability by integrating a native onion routing mechanism into the consensus protocol itself. PoS-CoPOR combines stake-weighted probabilistic leader election with an anonymization layer that conceals the network identity of the next block proposer. This approach prevents targeted DoS attacks on leaders before they produce a block, thus enhancing network resilience. We implemented and evaluated PoS-CoPOR, demonstrating its ability to achieve a throughput of up to 110 tx/s with 6 nodes, even with the overhead of the anonymization layer. The results show that native anonymization can provide robust DoS resistance with only a modest impact on performance, offering a solution to build secure and scalable PoS blockchains.

The proliferation of deepfake speech poses a significant threat to cybersecurity, from manipulating political speeches and impersonating public figures to spoofing voice biometric systems. The increasing sophistication of adversaries increases the necessity of deploying adaptive detection methods. Moreover, real-world incidents such as fraudulent financial transactions highlight the severity of the problem. Although numerous detectors have been developed, their evaluation remains difficult due to different methodologies and benchmark datasets, making direct comparisons impossible. This study presents a general and detailed framework for evaluating and comparing deepfake speech detectors. We further demonstrate the use of this framework to evaluate 40 state-of-the-art deepfake speech detectors under various conditions and data samples. We objectively compare these methods and identify the key attributes influencing performance the most. We also stress the issue of generalisation, as current detectors struggle to detect previously unseen deepfake speech samples or samples that have been modified. Finally, to strengthen the defence against synthetic audio content, we provide recommendations for improving the robustness of future detectors.

In this paper, we undertake a novel two-pronged investigation into the human recognition of deepfake speech, addressing critical gaps in existing research. First, we pioneer an evaluation of the impact of prior information on deepfake recognition, setting our work apart by simulating real-world attack scenarios where individuals are not informed in advance of deepfake exposure. This approach simulates the unpredictability of real-world deepfake attacks, providing unprecedented insights into human vulnerability under realistic conditions. Second, we introduce a novel metric to evaluate the quality of deepfake audio. This metric facilitates a deeper exploration into how the quality of deepfake speech influences human detection accuracy. By examining both the effect of prior knowledge about deepfakes and the role of deepfake speech quality, our research reveals the importance of these factors, contributes to understanding human vulnerability to deepfakes, and suggests measures to enhance human detection skills.

This paper explores the growing significance of vulnerability disclosure and bug bounty programs within the cybersecurity landscape, driven by regulatory changes in the European Union. The effectiveness of these programs relies heavily on the expertise of participants, presenting a challenge amid a shortage of skilled cybersecurity professionals, particularly in less sought-after sectors. To address this issue, the paper proposes a collaborative approach between academia and bug bounty issuers. By integrating bug bounty programs into cybersecurity courses, students gain practical skills and soft skills essential for bug hunting and cybersecurity work. The collaboration benefits both issuers, who gain manageable manpower, and students, who receive valuable hands-on experience. A pilot conducted during the current academic year yielded positive results, indicating the potential of this approach to address the demand for skilled cybersecurity professionals. The insights gained from the pilot inform future considerations and advancements in this collaborative model.

AI assistants for coding are on the rise. However one of the reasons developers and companies avoid harnessing their full potential is the questionable security of the generated code. This paper first reviews the current state-of-the-art and identifies areas for improvement on this issue. Then, we propose a systematic approach based on prompt-altering methods to achieve better code security of (even proprietary black-box) AI-based code generators such as GitHub Copilot, while minimizing the complexity of the application from the user point-of-view, the computational resources, and operational costs. In sum, we propose and evaluate three prompt altering methods: (1) scenario-specific, (2) iterative, and (3) general clause, while we discuss their combination. Contrary to the audit of code security, the latter two of the proposed methods require no expert knowledge from the user. We assess the effectiveness of the proposed methods on the GitHub Copilot using the OpenVPN project in realistic scenarios, and we demonstrate that the proposed methods reduce the number of insecure generated code samples by up to 16\% and increase the number of secure code by up to 8\%. Since our approach does not require access to the internals of the AI models, it can be in general applied to any AI-based code synthesizer, not only GitHub Copilot.