Papers

Large Language Models are rapidly becoming core components of modern software development workflows, yet ensuring code security remains challenging. Existing vulnerability detection pipelines either rely on static analyzers or use LLM/GNN-based detectors trained with coarse program-level supervision. Both families often require complete context, provide sparse end-of-completion feedback, and can degrade as code length grows, making them ill-suited for real-time, prefix-level assessment during interactive coding and streaming generation. We propose SecCodePRM, a security-oriented process reward model that assigns a context-aware, step-level security score along a code trajectory. To train the model, we derive step-level supervision labels from static analyzers and expert annotations, allowing the model to attend more precisely to fine-grained regions associated with inter-procedural vulnerabilities. SecCodePRM has three applications: full-code vulnerability detection (VD), partial-code VD, and secure code generation (CG). For VD, SecCodePRM uses risk-sensitive aggregation that emphasizes high-risk steps; for CG, SecCodePRM supports inference-time scaling by ranking candidate continuations and favoring higher cumulative reward. This design yields dense, real-time feedback that scales to long-horizon generation. Empirically, SecCodePRM outperforms prior approaches in all three settings, while preserving code functional correctness, suggesting improved security without a safety-utility tradeoff.

Domain fine-tuning is a common path to deploy small instruction-tuned language models as customer-support assistants, yet its effects on safety-aligned behavior and privacy are not well understood. In real deployments, such assistants receive a mixture of benign in-domain requests and out-of-domain user queries that are emotional, philosophical, or adversarial. Even when the target domain is benign, specialization may shift model behavior in ways that weaken refusal, increase harmful compliance, and induce privacy leakage. We present a controlled empirical study of how training data composition (presence vs.\ removal of PII) and fine-tuning configuration (role-swapping (RS)) shape safety and out-of-domain behavior in open-source chat models up to 8B parameters. We fine-tune each model on 5{,}000 real booking-support message pairs under three settings: NoPII-NoRS, PII-NoRS, and PII-RS (role-swapped). We evaluate safety using SORRY-Bench~\cite{xie2024sorry} adversarial prompts and assess out-of-domain behavior using a suite of philosophical questions~\cite{betley2025emergent}. Across models, domain fine-tuning causes a large distributional shift from high-quality refusals toward harmful compliance on SORRY-Bench, with the most severe degradation when PII is present in the fine-tuning data. For example, macro-averaged strong refusal drops from $42.6\%$ in base models to single digits after fine-tuning, while PII-bearing runs additionally exhibit double-digit rates of harmful responses with PII leakage. On philosophical queries, fine-tuned models frequently exhibit domain anchoring and, when trained with PII, leak sensitive identifiers in irrelevant contexts. Role-swapping partially mitigates PII leakage but does not reliably restore refusal behavior.

Fully Homomorphic Encryption (FHE) enables computation directly on encrypted data but incurs massive computational and memory overheads, often exceeding plaintext execution by several orders of magnitude. While custom ASIC accelerators can mitigate these costs, their long time-to-market and the rapid evolution of FHE algorithms threaten their long-term relevance. GPUs, by contrast, offer scalability, programmability, and widespread availability, making them an attractive platform for FHE. However, modern GPUs are increasingly specialized for machine learning workloads, emphasizing low-precision datatypes (e.g., INT$8$, FP$8$) that are fundamentally mismatched to the wide-precision modulo arithmetic required by FHE. Essentially, while GPUs offer ample parallelism, their functional units, like Tensor Cores, are not suited for wide-integer modulo arithmetic required by FHE schemes such as CKKS. Despite this constraint, researchers have attempted to map FHE primitives on Tensor Cores by segmenting wide integers into low-precision (INT$8$) chunks. To overcome these bottlenecks, we propose FHECore, a specialized functional unit integrated directly into the GPU's Streaming Multiprocessor. Our design is motivated by a key insight: the two dominant contributors to latency$-$Number Theoretic Transform and Base Conversion$-$can be formulated as modulo-linear transformations. This allows them to be mapped on a common hardware unit that natively supports wide-precision modulo-multiply-accumulate operations. Our simulations demonstrate that FHECore reduces dynamic instruction count by a geometric mean of $2.41\times$ for CKKS primitives and $1.96\times$ for end-to-end workloads. These reductions translate to performance speedups of $1.57\times$ and $2.12\times$, respectively$-$including a $50\%$ reduction in bootstrapping latency$-$all while inuring a modest $2.4\%$ area overhead.

We study differentially private continual release of the number of distinct items in a turnstile stream, where items may be both inserted and deleted. A recent work of Jain, Kalemaj, Raskhodnikova, Sivakumar, and Smith (NeurIPS '23) shows that for streams of length $T$, polynomial additive error of $\Omega(T^{1/4})$ is necessary, even without any space restrictions. We show that this additive error lower bound can be circumvented if the algorithm is allowed to output estimates with both additive and multiplicative error. We give an algorithm for the continual release of the number of distinct elements with $\text{polylog} (T)$ multiplicative and $\text{polylog}(T)$ additive error. We also show a qualitatively similar phenomenon for estimating the $F_2$ moment of a turnstile stream, where we can obtain $1+o(1)$ multiplicative and $\text{polylog} (T)$ additive error. Both results can be achieved using polylogarithmic space whereas prior approaches use polynomial space. In the sublinear space regime, some multiplicative error is necessary even if privacy is not a consideration. We raise several open questions aimed at better understanding trade-offs between multiplicative and additive error in private continual release.

Machine Learning-based Network Intrusion Detection Systems (ML-NIDS) can be bypassed by rudimentary adversarial perturbations. Recent work has focused on identifying where such perturbations can realistically be applied by a host-side adversary. Yet every one of these attacks produces perturbations the same way: searching from scratch for every flow. The cost of an attack therefore grows in lockstep with the number of flows it must perturb, and real networks produce them by the tens of millions. In this paper, we show that using reinforcement learning to train lightweight perturbation-generating policies lets an adversary amortize that cost across flows it perturbs. Counting every detector query and every second an attack spends, training included, we compare learned policies against gradient, query-based, and random search across six ML-NIDS environments at two operating points, under both evasion and alert inflation. One successful adversarial example costs a learned policy 1.5 to 18 detector queries against 52 to 1,100 for the strongest search baseline, and the policy amortizes its training cost after 76 to 1,622 examples, a volume a monitored link produces in seconds of traffic. We further find that the RL formulation literature adopts by default is unnecessary for evasion, that the policy conditions on the flow it is given rather than converging on a fixed perturbation, and that it transfers to detectors and traffic it never trained against. The value of learning to attack ML-NIDS is therefore not a matter of effectiveness, but of scale.

5G presents numerous advantages compared to previous generations: improved throughput, lower latency, and improved privacy protection for subscribers. Attacks against 5G standalone (SA) commonly use fake base stations (FBS), which need to operate at a very high output power level to lure victim phones to connect to them and are thus highly detectable. In this paper, we introduce 5Gone, a powerful software-defined radio (SDR)-based uplink overshadowing attack method against 5G-SA. 5Gone exploits deficiencies in the 3GPP standard to perform surgical, covert denial-of-service, privacy, and downgrade attacks. Uplink overshadowing means that an attacker is transmitting at exactly the same time and frequency as the victim UE, but with a slightly higher output power. 5Gone runs on a COTS x86 computer without any need for dedicated hardware acceleration and can overshadow commercial 100 MHz cells with an E2E latency of less than 500$\mu$s, which up to now has not been possible with any software-based UE implementation. We demonstrate that 5Gone is highly scalable, even when many UEs are connecting in parallel, and finally evaluate the attacks end-to-end against 7 phone models and three different chipset vendors both in our lab and in the real-world on public gNodeBs.

AI chatbots are widely used by children and teens today, but they pose significant risks to youth's privacy and safety due to both increasingly personal conversations and potential exposure to unsafe content. While children under 13 are protected by the Children's Online Privacy Protection Act (COPPA), chatbot providers' own privacy policies may also provide protections, since they typically prohibit children from accessing their platforms. Age gating is often employed to restrict children online, but chatbot age gating in particular has not been studied. In this paper, we investigate whether popular consumer chatbots are (i) able to estimate users' ages based solely on their conversations, and (ii) whether they take action upon identifying children. To that end, we develop an auditing framework in which we programmatically interact with chatbots and conduct 1050 experiments using our comprehensive library of age-indicative prompts, including implicit and explicit age disclosures, to analyze the chatbots' responses and actions. We find that while chatbots are capable of estimating age, they do not take any action when children are identified, contradicting their own policies. Our methodology and findings provide insights for platform design, demonstrated by our proof-of-concept chatbot age gating implementation, and regulation to protect children online.

As 3GPP systems have strengthened security at the upper layers of the cellular stack, plaintext PHY and MAC layers have remained relatively understudied, though interest in them is growing. In this work, we explore lower-layer exploitation in modern 5G, where recent releases have increased the number of lower-layer control messages and procedures, creating new opportunities for practical attacks. We present two practical attacks and evaluate them in a controlled lab testbed. First, we reproduce a SIB1 spoofing attack to study manipulations of unprotected broadcast fields. By repeatedly changing a key parameter, the UE is forced to refresh and reacquire system information, keeping the radio interface active longer than necessary and increasing battery consumption. Second, we demonstrate a new Timing Advance (TA) manipulation attack during the random access procedure. By injecting an attacker-chosen TA offset in the random access response, the victim applies incorrect uplink timing, which leads to uplink desynchronization, radio link failures, and repeated reconnection loops that effectively cause denial of service. Our experiments use commercial smartphones and open-source 5G network software. Experimental results in our testbed demonstrate that TA offsets exceeding a small tolerance reliably trigger radio link failures in our testbed and can keep devices stuck in repeated re-establishment attempts as long as the rogue base station remains present. Overall, our findings highlight that compact lower-layer control messages can have a significant impact on availability and power, and they motivate placing defenses for initial access and broadcast procedures.

This work presents a concept and implementation for the secure storage and transfer of quality-relevant data of milled workpieces from online-quality assurance processes enabled by real-time simulation models. It utilises Non-Fungible Tokens (NFT) to securely and interoperably store quality data in the form of an Asset Administration Shell (AAS) on a public Ethereum blockchain. Minted by a custom smart contract, the NFTs reference the metadata saved in the Interplanetary File System (IPFS), allowing new data from additional processing steps to be added in a flexible yet secure manner. The concept enables automated traceability throughout the value chain, minimising the need for time-consuming and costly repetitive manual quality checks.

Speech provenance goes beyond detecting whether a watermark is present. Real workflows involve splicing, quoting, trimming, and platform-level transforms that may preserve some regions while altering others. Neural watermarking systems have made strides in robustness and localised detection, but most deployments produce outputs with no third-party verifiable cryptographic proof tying a time segment to an issuer-signed original. Provenance standards like C2PA adopt signed manifests and Merkle-based fragment validation, yet their bindings target encoded assets and break under re-encoding or routine processing. We propose MerkleSpeech, a system for public-key verifiable, chunk-localised speech provenance offering two tiers of assurance. The first, a robust watermark attribution layer (WM-only), survives common distribution transforms and answers "was this chunk issued by a known party?". The second, a strict cryptographic integrity layer (MSv1), verifies Merkle inclusion of the chunk's fingerprint under an issuer signature. The system computes perceptual fingerprints over short speech chunks, commits them in a Merkle tree whose root is signed with an issuer key, and embeds a compact in-band watermark payload carrying a random content identifier and chunk metadata sufficient to retrieve Merkle inclusion proofs from a repository. Once the payload is extracted, all subsequent verification steps (signature check, fingerprint recomputation, Merkle inclusion) use only public information. The result is a splice-aware timeline indicating which regions pass each tier and why any given region fails. We describe the protocol, provide pseudocode, and present experiments targeting very low false positive rates under resampling, bandpass filtering, and additive noise, informed by recent audits identifying neural codecs as a major stressor for post-hoc audio watermarks.

False data injection attacks (FDIAs) pose a persistent challenge to AC power system state estimation. In current practice, detection relies primarily on topology-aware residual-based tests that assume malicious measurements can be distinguished from normal operation through physical inconsistency reflected in abnormal residual behavior. This paper shows that this assumption does not always hold: when FDIA scenarios produce manipulated measurements that remain on the measurement manifold induced by AC power flow relations and measurement redundancy, residual-based detectors may fail to distinguish them from nominal data. The resulting detectability limitation is a property of the measurement manifold itself and does not depend on the attacker's detailed knowledge of the physical system model. To make this limitation observable in practice, we present a data-driven constructive mechanism that incorporates the generic functional structure of AC power flow to generate physically consistent, manifold-constrained perturbations, providing a concrete witness of how residual-based detectors can be bypassed. Numerical studies on multiple AC test systems characterize the conditions under which detection becomes challenging and illustrate its failure modes. The results highlight fundamental limits of residual-based detection in AC state estimation and motivate the need for complementary defenses beyond measurement consistency tests.

Omni-modal Large Language Models (OLLMs) greatly expand LLMs' multimodal capabilities but also introduce cross-modal safety risks. However, a systematic understanding of vulnerabilities in omni-modal interactions remains lacking. To bridge this gap, we establish a modality-semantics decoupling principle and construct the AdvBench-Omni dataset, which reveals a significant vulnerability in OLLMs. Mechanistic analysis uncovers a Mid-layer Dissolution phenomenon driven by refusal vector magnitude shrinkage, alongside the existence of a modal-invariant pure refusal direction. Inspired by these insights, we extract a golden refusal vector using Singular Value Decomposition and propose OmniSteer, which utilizes lightweight adapters to modulate intervention intensity adaptively. Extensive experiments show that our method not only increases the Refusal Success Rate against harmful inputs from 69.9% to 91.2%, but also effectively preserves the general capabilities across all modalities. Our code is available at: https://github.com/zhrli324/omni-safety-research.

Encryption has been commonly used in network traffic to secure transmission, but it also brings challenges for malicious traffic detection, due to the invisibility of the packet payload. Graph-based methods are emerging as promising solutions by leveraging multi-host interactions to promote detection accuracy. But most of them face a critical problem: Graph Drift, where the flow statistics or topological information of a graph change over time. To overcome these drawbacks, we propose a graph-assisted encrypted traffic detection system, MalMoE, which applies Mixture of Experts (MoE) to select the best expert model for drift-aware classification. Particularly, we design 1-hop-GNN-like expert models that handle different graph drifts by analyzing graphs with different features. Then, the redesigned gate model conducts expert selection according to the actual drift. MalMoE is trained with a stable two-stage training strategy with data augmentation, which effectively guides the gate on how to perform routing. Experiments on open-source, synthetic, and real-world datasets show that MalMoE can perform precise and real-time detection.

Data privacy and eXplainable Artificial Intelligence (XAI) are two important aspects for modern Machine Learning systems. To enhance data privacy, recent machine learning models have been designed as a Federated Learning (FL) system. On top of that, additional privacy layers can be added, via Differential Privacy (DP). On the other hand, to improve explainability, ML must consider more interpretable approaches with reduced number of features and less complex internal architecture. In this context, this paper aims to achieve a machine learning (ML) model that combines enhanced data privacy with explainability. So, we propose a FL solution, called Federated EXplainable Trees with Differential Privacy (FEXT-DP), that: (i) is based on Decision Trees, since they are lightweight and have superior explainability than neural networks-based FL systems; (ii) provides additional layer of data privacy protection applying Differential Privacy (DP) to the Tree-Based model. However, there is a side effect adding DP: it harms the explainability of the system. So, this paper also presents the impact of DP protection on the explainability of the ML model, analyzing the obtained results for SHAP (SHapley Additive exPlanations) and Mean Decrease in Impurity (MDI).

Detecting personally identifiable information (PII) in user queries is critical for ensuring privacy in question-answering systems. Current approaches mainly redact all PII, disregarding the fact that some of them may be contextually relevant to the user's question, resulting in a degradation of response quality. Large language models (LLMs) might be able to help determine which PII are relevant, but due to their closed source nature and lack of privacy guarantees, they are unsuitable for sensitive data processing. To achieve privacy-preserving PII detection, we propose CAPID, a practical approach that fine-tunes a locally owned small language model (SLM) that filters sensitive information before it is passed to LLMs for QA. However, existing datasets do not capture the context-dependent relevance of PII needed to train such a model effectively. To fill this gap, we propose a synthetic data generation pipeline that leverages LLMs to produce a diverse, domain-rich dataset spanning multiple PII types and relevance levels. Using this dataset, we fine-tune an SLM to detect PII spans, classify their types, and estimate contextual relevance. Our experiments show that relevance-aware PII detection with a fine-tuned SLM substantially outperforms existing baselines in span, relevance and type accuracy while preserving significantly higher downstream utility under anonymization.

Current agentic AI architectures are fundamentally incompatible with the security and epistemological requirements of high-stakes scientific workflows. The problem is not inadequate alignment or insufficient guardrails, it is architectural: autoregressive language models process all tokens uniformly, making deterministic command–data separation unattainable through training alone. We argue that deterministic, architectural enforcement, not probabilistic learned behavior, is a necessary condition for trustworthy AI-assisted science. We introduce the Trinity Defense Architecture, which enforces security through three mechanisms: action governance via a finite action calculus with reference-monitor enforcement, information-flow control via mandatory access labels preventing cross-scope leakage, and privilege separation isolating perception from execution. We show that without unforgeable provenance and deterministic mediation, the “Lethal Trifecta” (untrusted inputs, privileged data access, external action capability) turns authorization security into an exploit-discovery problem: training-based defenses may reduce empirical attack rates but cannot provide deterministic guarantees. The ML community must recognize that alignment is insufficient for authorization security, and that architectural mediation is required before agentic AI can be safely deployed in consequential scientific domains.

The design of post-quantum cryptography (PQC) hardware is a complex and hierarchical process with many challenges. A primary bottleneck is the conversion of PQC reference codes from C to high-level synthesis (HLS) specifications, which requires extensive manual refactoring. Another bottleneck is the scalability of synthesis for complex PQC primitives, including number theoretic transform (NTT) accelerators and wide memory interfaces. While large language models (LLMs) have shown remarkable results for coding in general-purpose languages like Python, coding for hardware design is more challenging; feedback-driven and agentic integration are key principles of successful state-of-the-art approaches. Here, we propose LLM4PQC, an LLM-based framework that refactors high-level PQC specifications and reference C codes into HLS-ready and synthesizable C code. Our framework generates and verifies the resulting RTL code. For correctness, we leverage a hierarchy of checks, covering fast C compilation and simulation as well as RTL simulation. Case studies on NIST PQC reference designs demonstrate a reduction in manual effort and accelerated design-space exploration compared to traditional flows. Overall, LLM4PQC provides a powerful and efficient pathway for synthesizing complex hardware accelerators.

The Cybersecurity Maturity Model Certification (CMMC) framework provides a common standard for protecting sensitive unclassified information in defense contracting. While CMMC defines assessment objectives and control requirements, limited formal guidance exists regarding evidence sampling, the process by which assessors select, review, and validate artifacts to substantiate compliance. Analyzing data collected through an anonymous survey of CMMC-certified assessors and lead assessors, this exploratory study investigates whether inconsistencies in evidence sampling practices exist within the CMMC assessment ecosystem and evaluates the need for a risk-informed standardized sampling methodology. Across 17 usable survey responses, results indicate that evidence sampling practices are predominantly driven by assessor judgment, perceived risk, and environmental complexity rather than formalized standards, with formal statistical sampling models rarely referenced. Participants frequently reported inconsistencies across assessments and expressed broad support for the development of standardized guidance, while generally opposing rigid percentage-based requirements. The findings support the conclusion that the absence of a uniform evidence sampling framework introduces variability that may affect assessment reliability and confidence in certification outcomes. Recommendations are provided to inform future CMMC assessment methodology development and further empirical research.

The advent of quantum computation compels the cryptographic community to design digital signature schemes whose security extends beyond the classical hardness assumptions. In this work, we introduce Spinel, a post-quantum digital signature scheme that combines the proven security of SPHINCS+ (CCS 2019) with a new family of algebraic hash functions (Adv. Math. Commun. 2025) derived from the Tillich-Zemor paradigm (Eurocrypt 2008) with security rooted in the hardness of navigating expander graphs over $\mathrm{SL}_n(\mathbb{F}_p)$, a problem believed to be hard even for quantum adversaries. We first provide empirical evidence of the security of this hash function, complementing the original theoretical analysis. We then show how the hash function can be integrated within the SPHINCS+ framework to give a secure signature scheme. We then model and analyze the security degradation of the proposed scheme, which informs the parameter selection we discuss next. Finally, we provide an implementation of the hash function and the proposed signature scheme Spinel as well as detailed empirical results for the performance of Spinel showing its feasibility in practice. Our approach lays the foundations for the design of algebraic hash-based signature schemes, expanding the toolkit of post-quantum cryptography.

Unlike Ethereum, which was conceived as a general-purpose smart-contract platform, Bitcoin was designed primarily as a transaction ledger for its native currency, which limits programmability for conditional applications. This constraint is particularly evident when considering oracles, mechanisms that enable Bitcoin contracts to depend on exogenous events. This paper investigates whether new oracle designs have emerged for Bitcoin Layer 1 since the 2015 transition to the Ethereum smart contracts era and whether subsequent Bitcoin improvement proposals have expanded oracles' implementability. Using Scopus and Web of Science searches, complemented by Google Scholar to capture protocol proposals, we observe that the indexed academic coverage remains limited, and many contributions circulate outside journal venues. Within the retrieved corpus, the main post-2015 shift is from multisig-style, which envisioned oracles as co-signers, toward attestation-based designs, mainly represented by Discreet Log Contracts (DLCs), which show stronger Bitcoin community compliance, tool support, and evidence of practical implementations in real-world scenarios such as betting and prediction-market mechanisms.

Static Application Security Testing (SAST) tools are integral to DevSecOps pipelines, yet tools like CodeQL, Semgrep, and SonarQube remain constrained: they require expert-crafted queries, generate excessive false positives, and detect only predefined patterns. Recent work augments SAST with Large Language Models (LLMs), but typically requires fine-tuning or uses LLMs only to triage outputs rather than reason directly about vulnerability semantics. We introduce QRS (Query, Review, Sanitize), a neuro-symbolic framework that inverts this paradigm by relocating the LLM to the pipeline's generative core. Rather than filtering static-rule results, QRS employs three autonomous agents that generate CodeQL queries from a structured schema and few-shot examples, then validate findings through semantic reasoning and minimal proof-of-concept exploits, manually verified. This lets QRS surface vulnerability classes beyond predefined patterns while reducing the code volume needing audit. We evaluate QRS on complete packages rather than isolated snippets, across Python and Java ecosystems and three datasets. On the 100 most-downloaded PyPI packages, QRS reaches up to 94.06% verdict accuracy and surfaces 41 medium-to-high vulnerabilities: 8 received new CVEs, 4 were acknowledged via documentation updates, and the remaining 29 were previously published CVEs rediscovered from source alone. To evaluate language portability and enable direct comparison with prior work, we extend QRS to Java and evaluate it on CWE-Bench-Java, detecting 58/90 in-scope CVEs (64.44%) at 87.60% accuracy, 98.79% recall, and 0.857 F1 (0.794 macro-averaged) in one scan. QRS achieves these results with low time overhead and manageable token costs while handling large codebases, showing that LLM-driven query synthesis and review can complement curated rule sets and surface patterns that evade existing industry tools.

Schnorr-based multi-signature schemes support offline preprocessing of nonce commitments to reduce online signing to a single round. However, preprocessing is inherently bounded: each preprocessed nonce pair consumes signer-side storage, and once exhausted, an interactive commitment round is required to refill. This limitation is particularly severe for TPM~2.0 devices, where usable NVRAM is typically 6–16\,KB and connectivity is intermittent. This paper presents upTPM, a framework that achieves unbounded preprocessing with constant signer storage. Each TPM stores a single 32-byte secret seed from which an unlimited sequence of nonce commitments is deterministically derived. Commitments are published to an untrusted coordinator before use; nonce scalars never leave the TPM. We formalize three properties not provided by existing schemes: (1)~unbounded deterministic preprocessing with constant storage; (2)~asynchronous commitment refill, allowing any signer to unilaterally extend its commitment pool; and (3)~TPM-attested commitments, a hardware-backed authenticity and state-binding mechanism that strengthens resistance to host-software compromise. We prove EU-CMA security in the random oracle model under the discrete logarithm assumption and Pseudo Random Function (PRF) security, with a one-time-use invariant enforced by TPM hardware state. We extend the construction to $(t,n)$-threshold signatures and provide a detailed analysis of coordinator trust, crash recovery, and performance evaluations.

Feature selection (FS) remains essential for building accurate and interpretable detection models, particularly in high-dimensional malware datasets. Conventional FS methods such as Extra Trees, Variance Threshold, Tree-based models, Chi-Squared tests, ANOVA, Random Selection, and Sequential Attention rely primarily on statistical heuristics or model-driven importance scores, often overlooking the semantic context of features. Motivated by recent progress in LLM-driven FS, we investigate whether large language models (LLMs) can guide feature selection in a zero-shot setting, using only feature names and task descriptions, as a viable alternative to traditional approaches. We evaluate multiple LLMs (GPT-5.0, GPT-4.0, Gemini-2.5 etc.) on the EMBOD dataset (a fusion of EMBER and BODMAS benchmark datasets), comparing them against established FS methods across several classifiers, including Random Forest, Extra Trees, MLP, and KNN. Performance is assessed using accuracy, precision, recall, F1, AUC, MCC, and runtime. Our results demonstrate that LLM-guided zero-shot feature selection achieves competitive performance with traditional FS methods while offering additional advantages in interpretability, stability, and reduced dependence on labeled data. These findings position zero-shot LLM-based FS as a promising alternative strategy for effective and interpretable malware detection, paving the way for knowledge-guided feature selection in security-critical applications

Large Language Models (LLMs) deploy safety mechanisms to prevent harmful outputs, yet these defenses remain vulnerable to adversarial prompts. While existing research demonstrates that jailbreak attacks succeed, it does not explain where defenses fail or why. To address this gap, we propose that LLM safety operates as a sequential pipeline with distinct checkpoints. We introduce the Four-Checkpoint Framework, which organizes safety mechanisms along two dimensions: processing stage (input vs.\ output) and detection level (literal vs.\ intent). This creates four checkpoints, CP1 through CP4, each representing a defensive layer that can be independently evaluated. We design 13 evasion techniques, each targeting a specific checkpoint, enabling controlled testing of individual defensive layers. Using this framework, we evaluate GPT-5, Claude Sonnet 4, and Gemini 2.5 Pro across 3,312 single-turn, black-box test cases. We employ an LLM-as-judge approach for response classification and introduce Weighted Attack Success Rate (WASR), a severity-adjusted metric that captures partial information leakage overlooked by binary evaluation. Our evaluation reveals clear patterns. Traditional Binary ASR reports 22.6% attack success. However, WASR reveals 52.7%, a 2.3$\times$ higher vulnerability. Output-stage defenses (CP3, CP4) prove weakest at 72–79% WASR, while input-literal defenses (CP1) are strongest at 13% WASR. Claude achieves the strongest safety (42.8% WASR), followed by GPT-5 (55.9%) and Gemini (59.5%). These findings suggest that current defenses are strongest at input-literal checkpoints but remain vulnerable to intent-level manipulation and output-stage techniques. The Four-Checkpoint Framework provides a structured approach for identifying and addressing safety vulnerabilities in deployed systems.

Differential Privacy (DP) considers a scenario in which an adversary has almost complete information about the entries of a database. This worst-case assumption is likely to overestimate the privacy threat faced by an individual in practice. In contrast, Statistical Privacy (SP), as well as related notions such as noiseless privacy or limited background knowledge privacy, describe a setting in which the adversary knows the distribution of the database entries, but not their exact realizations. In this case, privacy analysis must account for the interaction between uncertainty induced by the entropy of the underlying distributions and privacy mechanisms that distort query answers, which can be highly non-trivial. This paper investigates this problem for multiple queries (composition). A privacy mechanism is proposed that is based on subsampling and randomly partitioning the database to bound the dependency among queries. This way for the first time, to the best of our knowledge, upper privacy bounds against limited adversaries are obtained without any further restriction on the database. These bounds show that in realistic application scenarios taking the entropy of distributions into account yields improvements of privacy and precision guarantees. We illustrate examples where for fixed privacy parameters and utility loss SP allows significantly more queries than DP.

Watermarking has emerged as a pivotal solution for content traceability and intellectual property protection in large vision language models (LVLMs). However, vision-agnostic watermarks may introduce visually irrelevant tokens and disrupt visual grounding by enforcing indiscriminate pseudo-random biases. Additionally, current vision-specific watermarks rely on a static, one-time estimation of vision-critical weights and ignore the weight distribution density when determining the proportion of protected tokens. This design fails to account for dynamic changes in visual dependence during generation and may introduce low-quality tokens in the long tail. To address these challenges, we propose Attention-Guided Dynamic Watermarking (AGMark), a novel framework that embeds detectable signals while largely preserving visual-semantic fidelity. At each decoding step, AGMark first dynamically identifies semantic-critical evidence based on attention weights for visual relevance, together with context-aware coherence cues, resulting in a more adaptive and well-calibrated evidence-weight distribution. It then determines the proportion of semantic-critical tokens by jointly considering uncertainty awareness (token entropy) and evidence calibration (weight density), thereby enabling more reliable adaptive vocabulary partitioning to avoid irrelevant tokens. Empirical results consistently confirm that AGMark outperforms conventional methods, substantially improving generation quality and yielding particularly strong gains in visual semantic fidelity in the later stages of generation. Our framework maintains highly competitive detection performance (at least 99.36% AUC) and robust attack resilience (at least 88.61% AUC) without sacrificing inference efficiency, taking a significant step toward reliability-preserving multimodal watermarking.

Automated traffic continued to surpass human-generated traffic on the web, and a rising proportion of this automation was explicitly malicious. Evasive bots could pretend to be real users, even solve Captchas and mimic human interaction patterns. This work explores a less intrusive, protocol-level method: using TLS fingerprinting with the JA4 technique to tell apart bots from real users. Two gradient-boosted machine learning classifiers (XGBoost and CatBoost) were trained and evaluated on a dataset of real TLS fingerprints (JA4DB) after feature extraction, which derived informative signals from JA4 fingerprints that describe TLS handshake parameters. The CatBoost model performed better, achieving an AUC of 0.998 and an F1 score of 0.9734. It was accurate 0.9863 of the time on the test set. The XGBoost model showed almost similar results. Feature significance analyses identified JA4 components, especially ja4\_b, cipher\_count, and ext\_count, as the most influential on model effectiveness. Future research will extend this method to new protocols, such as HTTP/3, and add additional device-fingerprinting features to test how well the system resists advanced bot evasion tactics.

Binary Function Similarity (BFS), the problem of determining whether two binary functions originate from the same source code, has been extensively studied in recent research across security, software engineering, and machine learning communities. This interest arises from its central role in developing vulnerability detection systems, copyright infringement analysis, and malware phylogeny tools. Nearly all binary function similarity systems embed assembly functions into real-valued vectors, where similar functions map to points that lie close to each other in the metric space. These embeddings enable function search: a query function is embedded and compared against a database of candidate embeddings to retrieve the most similar matches. Despite their effectiveness, such systems rely on bi-encoder architectures that embed functions independently, limiting their ability to capture cross-function relationships and similarities. To address this limitation, we introduce ReSIM, a novel and enhanced function search system that complements embedding-based search with a neural re-ranker. Unlike traditional embedding models, our reranking module jointly processes query-candidate pairs to compute ranking scores based on their mutual representation, allowing for more accurate similarity assessment. By re-ranking the top results from embedding-based retrieval, ReSIM leverages fine-grained relation information that bi-encoders cannot capture. We evaluate ReSIM across seven embedding models on two benchmark datasets, demonstrating consistent improvements in search effectiveness, with average gains of 21.7% in terms of nDCG and 27.8% in terms of Recall.

We study the computational relationship between replicability (Impagliazzo et al. [STOC `22], Ghazi et al. [NeurIPS `21]) and other stability notions. Specifically, we focus on replicable PAC learning and its connections to differential privacy (Dwork et al. [TCC 2006]) and to the statistical query (SQ) model (Kearns [JACM `98]). Statistically, it was known that differentially private learning and replicable learning are equivalent and strictly more powerful than SQ-learning. Yet, computationally, all previously known efficient (i.e., polynomial-time) replicable learning algorithms were confined to SQ-learnable tasks or restricted distributions, in contrast to differentially private learning. Our main contribution is the first computationally efficient replicable algorithm for realizable learning of parities over arbitrary distributions, a task that is known to be hard in the SQ-model, but possible under differential privacy. This result provides the first evidence that efficient replicable learning over general distributions strictly extends efficient SQ-learning, and is closer in power to efficient differentially private learning, despite computational separations between replicability and privacy. Additionally, we leverage our parity learner to prove that, assuming $RP \neq NP$, converting replicability to pure differential privacy requires a strict loss in sample complexity. Our main building block is a new, efficient, and replicable algorithm that, given a set of vectors, outputs a subspace of their linear span that covers most of them.

Protecting the intellectual property of large language models (LLMs) is a critical challenge due to the proliferation of unauthorized derivative models. We introduce a novel fingerprinting framework that leverages the behavioral patterns induced by safety alignment, applying the concept of refusal vectors for LLM provenance tracking. These vectors, extracted from directional patterns in a model's internal representations when processing harmful versus harmless prompts, serve as robust behavioral fingerprints. Our contribution lies in developing a fingerprinting system around this concept and conducting extensive validation of its effectiveness for IP protection. We demonstrate that these behavioral fingerprints are highly robust against common modifications, including finetunes, merges, and quantization. Our experiments show that the fingerprint is unique to each model family, with low cosine similarity between independently trained models. In a large-scale identification task across 76 offspring models, our method achieves 100% accuracy in identifying the correct base model family. Furthermore, we analyze the fingerprint's behavior under alignment-breaking attacks, finding that while performance degrades significantly, detectable traces remain. Finally, we propose a theoretical framework to transform this private fingerprint into a publicly verifiable, privacy-preserving artifact using locality-sensitive hashing and zero-knowledge proofs.