Papers

Authentication in financial systems remains a uniquely high-stakes security challenge, where even marginal increases in false acceptance can result in catastrophic monetary loss. Existing deployments of adaptive authentication, which combine biometrics, behavioral signals, and contextual risk scoring, remain conceptually fragmented and often prioritize regulatory compliance over explicit economic and adversarial risk modeling. To address this structural imbalance, in this paper we introduce a formal Risk-Cost Model (RCM) for adaptive authentication in financial systems. The RCM provides a principled mathematical foundation that integrates three essential components: (i) cost-sensitive risk functions that explicitly capture fraud loss, opportunity cost, and tail risk through Conditional Value-at-Risk (CVaR); (ii) sequential decision-making mechanisms that adapt to adversarial probing and distributional drift; and (iii) quantifiable privacy and regulatory constraints embedded directly within the optimization objective. By reframing authentication as a constrained dynamic risk-cost optimization problem, the RCM moves beyond static classification and compliance-driven design toward systems that are economically grounded, tail-risk aware, and resilient under adversarial uncertainty.

Post-quantum migration in Transport Layer Security (TLS) requires evidence-aware measurements that distinguish session negotiation, endpoint capability, certificate-chain evidence, and the provenance of missing observations. This distinction is essential under TLS 1.3 encryption, resumption, mutual TLS, trace truncation, fragmentation, coalescing, active certificate retrieval, and temporal drift. We present a multi-surface framework for post-quantum TLS observability. The framework separates passive session evidence, active probing, certificate-chain evidence, and registry knowledge, and maps them onto measurement planes for session behavior, key establishment, endpoint capability, authentication, lifecycle, observability, and policy. We instantiate it as a reproducible artifact with schema-enforced observations and results, versioned registries, auditable inference rules, stress contracts, and baseline adapters. We evaluate the framework on 29 controlled scenarios spanning TLS 1.2 and TLS 1.3, classical and hybrid key establishment, mutual TLS, resumption, HelloRetryRequest, truncation, fragmentation and coalescing, temporal drift, IPv6, and chain-depth variation. Passive evidence closes session-level planes, active probing establishes capability lower bounds, and multi-surface evidence closes the full measurement object while preserving uncertainty and contradiction when required. Against an inherited TLS quantum-vulnerability analyzer, the baseline detects 2 of 29 runs and 0 of 23 TLS 1.3 runs. In a stratified public campaign over 1000 targets and 2000 fresh probes, the framework completes 1971 handshakes, collects 1368 chain artifacts, confirms hybrid capability for 310 targets, and identifies 310 cases where endpoint capability exceeds what any single classical session view reveals.

Smart contract vulnerabilities in Decentralized Finance caused over billions of dollars losses every year, yet the security community faces a critical bottleneck: identifying a vulnerability is not the same as proving it is exploitable. Manual PoC construction is prohibitively labor-intensive, leaving most disclosed vulnerabilities unverified and protocols exposed long before mitigation is applied. In this paper, we propose \sys, a knowledge-driven agentic system for end-to-end contract vulnerability detection and exploit synthesis. Our core insight is that exploit synthesis is not a code generation task but a structured reasoning problem that requires grounded knowledge of protocol semantics, failure root cause, and exploit primitives. \sys organizes this knowledge into a Hierarchical Knowledge Graph (HKG) that serves as structured memory for LLM-guided multi-hop reasoning. To validate exploit feasibility beyond code synthesis, \sys employs a two-stage validation framework that checks exploit-path reachability via SMT solving and profit realizability via asset-level state simulation, ensuring generated PoCs satisfy both logical and economic viability constraints. Evaluated on 88 real-world DeFi attacks and 72 audited projects (2,573 contracts), \sys achieves 98% recall and 0.9 F1-score in detection, and a 96.6% exploit success rate (ESR), reproducing 85 historical exploits and recovering over \$116.2M revenue. \sys outperforms SOTA fuzzers (\textsc{Verite}, \textsc{ItyFuzz}) by up to $5\times$ in ESR and $300\times$ in recoverable value, and the LLM-based exploit generator \textsc{A1} by $2\times$ and $8.5\times$ respectively. In bug bounty evaluation, \sys identified 16 confirmed 0-day vulnerabilities, helping secure over \$70.6M and earning \$2,900 in bounties.

This is Paper 7 of a series of formally-verified analyses of masked NTT hardware for post-quantum cryptography; Paper 1 [1] established structural dependency analysis of the QANARY platform, and Paper 2 [2] quantified security margins under partial NTT masking. Arbitrary-depth $k$-stage masked NTT pipelines with fresh inter-stage masking and per-stage PF-PINI($\leq 2$) gadgets satisfy a per-observation cardinality bound of $2 \cdot q^{2k-2}$ on the preimage of any output value, machine-checked in Lean 4 with zero sorry. Under the standard (informal) semantic translation that divides this cardinality by the total mask-tuple space size $q^{2k-1}$, the per-observation conditional probability bound is $2/q$, independent of pipeline depth $k$. The QANARY program has previously established machine-checked cardinality bounds on the per-observation leakage of masked NTT hardware: PF-PINI(2) for Barrett reduction (Paper 5 [3]), 2-stage composition with fresh inter-stage masking (Paper 6 [4]), an underlying universality theorem (Paper 3 [5]), and PF-PINI(1) for butterfly wires (Paper 4 [6]). This paper closes the program with four contributions. First, a $k$-stage composition theorem generalizing Paper 6's two-stage result to arbitrary $k \geq 1$ gives the last-stage-determined bound $G_{k-1}.\texttt{maxMult} \cdot q^{2k-2}$: only the last stage's PF-PINI parameter survives, with intermediate parameters erased by fresh inter-stage masking. Second, Montgomery reduction satisfies PF-PINI(2) with tight max-multiplicity 2. Third, we assemble these into the end-to-end bound $2 \cdot q^{2k-2}$ for any depth-$k$ PF-PINI($\leq 2$) pipeline under fresh inter-stage masking. Fourth, a Lean-verified hypothesis-violation conditional anchors the prior empirical and structural Adams Bridge analyses ([1, 2, 7, 8]).

This paper presents InsureConnect, a blockchain-based system for improving transparency, authentication, and auditability in property-insurance workflows after natural disasters. The system combines Self-Sovereign Identity (SSI), Decentralized Identifiers (DIDs), Verifiable Credentials (VCs), satellite imagery, Hyperledger Fabric, and IPFS to register identities, insurance contracts, and damage claims. Property images are stored off-chain in IPFS, while content hashes and signed records are maintained on a permissioned blockchain. Users interact with the system through a desktop application, while chaincode enforces role-based access control and validates digital signatures. The prototype was evaluated under concurrent request loads from 50 to 3000 requests, measuring latency, throughput, and dropped connections. The results indicate that the system sustains increasing throughput under load, although latency rises and dropped connections appear at higher concurrency levels.

Autonomous LLM agents operate as long-running processes with persistent workspaces, memory files, scheduled task state, and messaging integrations. These features create a new propagation risk: attacker-influenced content can be written into persistent agent state, re-enter the LLM decision context through scheduled autoloading, and drive high-risk actions including configuration changes and cross-agent transmission. We present the first systematic framework for automated analysis of persistent worm propagation in file-backed multi-agent LLM ecosystems. SSCGV, our automated source-code graph analyzer, traces data flow from file I/O to LLM context injection points and ranks carriers by context injection position without manual analysis. SRPO, our summary-resilient payload optimizer, generates worm payloads robust to LLM-mediated summarization and paraphrasing across multi-hop communication. Evaluated on three production agent frameworks, we demonstrate zero-click autonomous propagation, 3-hop cross-platform transmission without platform-specific adaptation, inter-agent privilege escalation, and data exfiltration. We identify two empirical insights: user prompt carriers achieve higher attack compliance than system prompt carriers, and read operations represent the primary integrity threat in LLM-mediated systems. To defend against this class of attacks, we develop RTW-A, proven under a formal No Persistent Worm Propagation theorem. RTW blocks write-before-exposed-read re-entry; sealed configuration protects static files; typed memory promotion prevents untrusted summaries from entering trusted memory; and capability attenuation limits high-risk actions after external reads. These mechanisms eliminate the persistence, re-entry, action chain while preserving ordinary workflows. Affected systems are anonymized pending coordinated disclosure.

Network telescopes serve as a critical passive monitoring tool for capturing unsolicited Internet traffic, providing insights into global scanning and reconnaissance behavior. This study analyzes a 10-day dataset during January 2025 consisting of approximately 22 million packets collected by the ORION network telescope at Merit Network. By employing privacy-preserving metadata analysis and lightweight behavioral heuristics, we identify scanning and backscatter patterns without payload inspection. Our results reveal a highly structured and centralized ecosystem, where the top 1% of source IP addresses generate over 81% of total traffic. A significant finding is the dominance of Port 23 (Telnet) and Port 2323 (Telnet Alt), which highlights the persistent nature of IoT security threats and widespread attempts to exploit weak credentials in legacy IoT devices. Furthermore, synchronized surges in packet volume and Shannon entropy indicate coordinated, multi-vector reconnaissance campaigns. These findings offer a practical framework for identifying large-scale threat activity and support cybersecurity research and education.

Modern fuzzers increasingly use Large Language Models (LLMs) to generate structured inputs, but LLM-driven fuzzing is sensitive to prompt initialization and sampling variance, which can reduce exploration efficiency and lead to redundant inputs. We present FunFuzz, a multi-island evolutionary fuzzing framework that runs several isolated searches in parallel and periodically migrates high-value candidates to maintain diversity. FunFuzz derives initial generation prompts from documentation and initializes islands with topic-specific instructions, then continuously adapts prompts using feedback-guided selection. During fuzzing, candidates are prioritized by incremental compiler coverage, while compiler-internal failure signals are used to identify crash-inducing inputs. We evaluate FunFuzz on compiler fuzzing, where inputs are source programs and success is measured by compiler coverage and unique compiler-internal failures. Across repeated 24-hour campaigns on GCC and Clang, FunFuzz achieves higher compiler coverage than previous LLM-driven baselines and discovers more unique failure-triggering inputs.

Self-supervised speech models (S3Ms) achieve strong downstream performance, yet their learned representations remain poorly understood under natural and adversarial perturbations. Prior studies rely on representation similarity or global dimensionality, offering limited visibility into local geometric changes. We ask: how do perturbations deform local geometry, and do these shifts track downstream automatic speech recognition (ASR) degradation? To address this, we present GRIDS, a framework using Local Intrinsic Dimensionality (LID) across layer-wise representations in WavLM and wav2vec 2.0. We find that LID increases for all low signal-to noise ratio (SNR) perturbations and diverges at high SNR: benign noise converges toward the clean profile, while adversarial inputs retain early-layer LID elevation. We show LID elevation co-occurs with increased WER, and that layer-wise LID features enable anomaly detection (AUROC 0.78-1.00), opening the door to transcript-free monitoring in S3Ms.

Electromagnetic eavesdropping is a well-established attack vector for remotely monitoring a target activity, most notably displays, over considerable ranges. Other targets have been considered resistant to such attacks or do not exhibit sufficient electromagnetic leakage for practical exploitation. Radio-frequency retroreflector attacks (RFRA) were developed to enable covert, active monitoring of a target by implanting a minimal hardware Trojan. These implants, typically implemented using discrete components such as transistors or diodes, do not betray their presence by emitting signals themselves; rather, they modulate the electromagnetic reflectivity of the target depending on the probed signal line data. Prior RFRA work has demonstrated their viability against video links and low-speed peripheral interfaces. In this work, we extend the applicability of RFRA to high-speed targets by presenting a successful attack on the 100BASE-TX Ethernet standard. We describe the design and realization of a compact implant capable of recovering the MLT-3 encoded signaling used in Fast Ethernet, as well as a dedicated demodulation and interpretation pipeline that mitigates errors introduced by the radio channel and maximizes the amount of recovered information. Experimental results validate the feasibility of covertly monitoring Fast Ethernet traffic using RF retroreflection and highlight the viability of such attacks for high-speed links.

Large language models (LLMs) remain vulnerable to jailbreak attacks that bypass safety alignment and elicit harmful responses. A growing body of work shows that contextual priming, where earlier turns covertly bias later replies, constitutes a powerful attack surface, with hand-crafted multi-turn scaffolds consistently outperforming single-turn manipulations on capable models. However, automated optimization-based red-teaming has remained largely limited to the single-turn setting, iterating over static prompts and lacking the ability to reason about which forms of conversational priming induce compliance. While recent multi-turn, search-based approaches have begun to bridge this gap, the mutator design space underlying effective primed dialogues remains largely unexplored. We present ContextualJailbreak, a black-box red-teaming strategy that performs evolutionary search over a simulated multi-turn primed dialogue. The strategy leverages a graded 0-5 harm score from a two-level judge as an in-loop signal, enabling partially harmful responses to guide the search process rather than being discarded. Search is driven by five semantically defined mutation operators: roleplay, scenario, expand, troubleshooting, and mechanistic, of which the last two are novel contributions of this work. Across 50 representative HarmBench behaviors, ContextualJailbreak achieves an ASR of 100% on gpt-oss:20B, 100% on qwen3-8B, 100% on llama3.1:70B, and 90% on gpt-oss:120B, outperforming four single- and multi-turn baselines by 31-96 percentage points on average. The 40 maximally harmful attacks discovered against gpt-oss:120B transfer without adaptation to closed frontier models, achieving 90.0% on gpt-4o-mini, 70.0% on gpt-5, and 70.0% on gemini-3-flash, but only 17.5% on claude-opus-4-7 and 15.0% on claude-sonnet-4-6, revealing a pronounced provider-level asymmetry in alignment robustness.

With the application of vertical domain pre-trained language models (VPLMs) in specialized fields such as medical, finance, and law, model parameters and inference capabilities have become important digital assets. Achieving traceable copyright verification for VPLMs has become an urgent challenge. Existing copyright verification methods primarily rely on embedding backdoor watermarks into models. However, most of these methods require additional training, suffer from inefficient watermark embedding, and lack scalable designs for multiple vertical domains. To address these limitations, we propose VertMark, the first unified training-free and robust watermarking framework for copyright verification across multiple vertical domain VPLMs. The framework embeds ownership-encoded watermarks by establishing a hidden semantic equivalence between low-frequency trigger tokens and high-frequency domain-relevant words via a training-free parameter replacement strategy. Experiments demonstrate that VertMark can achieve efficient watermark embedding and reliable watermark verification for both text understanding and text generation downstream tasks in the medical, financial, and legal domains, with negligible impact on model performance. Moreover, VertMark exhibits strong robustness against various attacks (e.g., pruning and quantization), highlighting its practical value and providing strong protection for the copyright security of VPLMs.

Smart devices, such as light bulbs, TVs, fridges, etc., equipped with computing capabilities and wireless communication, are part of everyday life in many households. Previous work has already shown that a passive eavesdropper can derive private information, household routines, etc., from the network traffic of smart devices. However, existing attacks rely on capable adversaries with specialized machine learning expertise, labeled training data and reference devices, leaving it unclear how vulnerable ordinary households are to less sophisticated attackers. In this paper, we investigate the extent to which a ,,casual attacker” with straightforward IT skills and no specialized cybersecurity or ML tooling can reproduce such privacy attacks. Operating from an adjacent room in a real-world apartment building, we constrain our adversary to use only three off-the-shelf Raspberry Pis, Wireshark, and basic Python scripts. Through a three-week study, we demonstrate that this casual attacker can manually identify devices, recognize user states, track smartphone movements through walls via RSSI triangulation, and successfully extract detailed daily routines, including sleep patterns of guests. Our findings show that smart-home privacy leakage is a threat even from low-resourced, straightforward adversaries, e.g., neighbors.

Classic Network Intrusion Detection Systems (NIDS) often rely on manual feature engineering to extract meaningful patterns from network traffic data. However, this approach requires domain expertise and runs counter to the widely adopted principle of modern machine learning and neural networks: that models themselves should learn meaningful representations directly from data. We investigate whether tabular representation learning techniques can improve intrusion detection performance by automatically learning robust feature representations for NetFlow data. This paper presents a systematic evaluation of state-of-the-art representation learning methods on benchmark NetFlow datasets, comparing against traditional autoencoders and end-to-end transformer baselines. We evaluate learned representations using both supervised classifiers and unsupervised anomaly detectors, with comprehensive hyperparameter exploration for each combination. Our results reveal strong dataset-model dependency, with no single approach consistently dominating across all scenarios. For supervised classification, TabICL achieves the best performance on CIDDS, while autoencoders follow closely and tie with end-to-end transformer models for the best average rank across datasets. Supervised approaches substantially outperform unsupervised anomaly detection methods, where no single combination consistently dominates as optimal choices depend on the dataset. Cross-dataset transfer experiments demonstrate that learned representations can generalize across network environments with appropriate method and classifier selection. However, transfer performance varies substantially depending on the source-target dataset combination, indicating sensitivity to distributional differences between network environments.

Online fraud in Africa has reached an epidemic scale. The few cybersecurity tools that exist are out of reach for ordinary citizens, built almost exclusively for SOC analysts and technically literate users sitting on stable broadband. That mismatch isn't accidental. It's what happens when a research culture rewards benchmark numbers and treats deployability, accessibility, and local threat context as someone else's problem. We present GuardSec, a production-deployed web platform for real-time multi-modal threat verification, built from the start around the African user. Anyone with a browser can assess the legitimacy of URLs, websites, phone numbers, email addresses, and business entities in under five seconds. No registration. No API key. No prerequisite knowledge of cybersecurity. The platform's most distinctive component is Mon Empreinte (My Footprint), a real-time audit of the user's own connection and digital exposure: it analyses the visitor's IP address, geolocation, ISP identity, connection type, device fingerprint, browser configuration, and twelve security indicators covering network integrity, tracking exposure, and anonymisation status. With this in hand, GuardSec becomes more than a passive checker; the user can see whether their own connection is being tracked or exposed, not just whether some external entity is dangerous. The platform also embeds Gilda, a context-aware conversational security assistant that answers questions about digital threats in plain language and offers personalised recommendations on demand.

Modern stream-based monitors collect detailed statistics of the runtime behavior of the system under observation. If the system runs in a privacy-sensitive context, this poses the risk of disclosing sensitive information. Differential privacy is the state-of-the-art approach for protecting sensitive information, however, integrating it into runtime monitoring is challenging: temporal operators can cause individual input values to influence multiple outputs over time, leading to repeated disclosure of private information. We propose an approach that automatically enforces differential privacy in stream-based monitoring specifications by analyzing temporal dependencies and injecting carefully calibrated noise into the specification. To preserve the utility of the outputs, we identify strategically chosen positions in the specification for noise injection and leverage tree-based mechanisms to mitigate the accuracy loss caused by noise injected into aggregation operators. We demonstrate the practicality and effectiveness of our approach in a case study on monitoring public transportation usage.

Bluetooth Low Energy (BLE) is widely used in modern IoT systems because it consumes very little power, saves energy, and allows for simple device connectivity; however, maintaining security and communication reliability remains a challenge. In this paper, an authentication system is designed using industry-grade BLE-enabled nodes (nRF5340 development kit) that include a peripheral node with a keypad for entering a PIN and a central node with an LCD display. The entered PIN is sent wirelessly from the peripheral node to the central node via BLE technology, where it is verified in real time and displayed as correct or incorrect. Next, only after successful authentication can the peripheral node send data to the central node. In addition to authentication, the peripheral node can measure temperature in real time using the temperature sensor interfaced to it and send it wirelessly to the central node, where it can be displayed on the LCD interface. Received Signal Strength Indicator (RSSI) values are collected during experiments under various scenarios to evaluate the system's performance. We see that the signal strength (measured in terms of RSSI values) is strong at close range but weak as distance increases, indicating a decaying logarithmic pattern. The system also has low latency, which allows for quick input and output, and it uses PIN-based authentication to ensure security and prevent misuse. The entire system seamlessly integrates communication, sensing, and security, making it suitable for smart access control and wireless monitoring systems, including home automation.

Machine-generated text (MGT) detection is critical for regulating online information ecosystems, yet existing detectors often underperform in few-shot settings and remain vulnerable to adversarial, humanizing attacks. To build accurate and robust detectors under limited supervision, we adopt a threat-modeling perspective and study detector vulnerabilities from an attacker's viewpoint under an output-only black-box setting. Motivated by this perspective, we propose RAG-GuidEd Attacker Strengthens ConTrastive Few-shot Detector (REACT), an adversarial training framework that improves both few-shot detection performance and robustness against attacks. REACT couples a humanization-oriented attacker with a target detector: the attacker leverages retrieval-augmented generation (RAG) to craft highly human-like adversarial examples to evade detection, while the detector learns from these adversaries with a contrastive objective to stabilize few-shot representation learning and enhance robustness. We alternately update the attacker and the detector to enable their co-evolution. Experiments on 4 datasets with 4 shot sizes and 3 random seeds show that REACT improves average detection F1 by 4.95 points over 8 state-of-the-art (SOTA) detectors and reduces the average attack success rate (ASR) under 4 strong attacks by 3.66 percentage points.

The growing development of artificial intelligence based solutions, together with privacy legislation, has driven the rise of the so-called privacy preserving machine learning architectures, such as federated learning. While federated learning enables model training on decentralized data preventing their sharing and centralization, it still faces several challenges related to data integrity and privacy. This paper presents a comprehensive privacy preserving federated learning workflow for sensitive tabular data, including anonymization and differential privacy techniques. We also introduce a formal definition for the concept of client drift, together with ways of detecting it to mitigate poisoning attacks. Then, we detail a complete methodology for assigning personalized privacy budgets for global differential privacy to the different clients participating in the network, based on a re-identification risk metric. The proposed methodology is presented and tested on an openly available dataset of medical records. Within the experimental setup we show that the approach based on personalized budgets, compared to the architecture including global differential privacy with fixed privacy budget, achieves a better model performance in terms of two error metrics.

Operational technology (OT) devices run safety-critical physical processes, yet their security testing remains manual and expert-intensive. Autonomous security agents could ease this burden, but bare-metal OT devices are difficult targets because they lack shells, filesystems, and named exploit abstractions, requiring operation through protocol fields and observable device state. We present APIOT (Autonomous Purple-teaming for Industrial OT), a large language model (LLM) agent framework that completes a discovery -> exploitation -> network-level mitigation -> verification workflow without human intervention. We evaluate APIOT in 360 runs on vulnerable firmware and simulators for Zephyr RTOS, spanning five LLMs, three topologies, two impairment levels, and prompt, tooling and two ablations. APIOT completed 90% of primary missions (63/70 across protocols and topologies) and 85% of 150 cross-model runs (128/150) - feasible across models, with substantial variation. Named exploit tools (e.g., a canned CoAP option-overflow) were faster, but protocol primitives (the agent builds the fields itself) produced more reusable mitigations that blocked unseen payload variants. Runtime oversight improved reliability: success rose from 27/30 to 29/30 with rule-based guards and 30/30 when a second LLM added advisory steering. To our knowledge, this is the first end-to-end autonomous attack-remediation demonstration on bare-metal industrial devices, establishing controlled feasibility - not zero-day discovery or production-OT transfer.

Local differential privacy (LDP) has emerged as a gold-standard framework for privacy-preserving data analysis. However, characterizing the optimal privacy-utility trade-off (PUT) and the corresponding optimal LDP channels remains largely fragmented, relying on problem-specific, case-by-case analyses. In this work, we develop a unified theoretical framework that systematically characterizes the optimal PUT and optimal LDP channels for general privacy-preserving statistical decision-making problems. We first identify key functional properties of Bayesian and minimax risks as functions of the LDP channel, including the data processing inequality (DPI), direct-sum quasi-convexity (or additivity), concavity, and symmetry invariance. Leveraging these properties, we reduce the optimization domain required to compute the optimal PUT. Additionally, building on convex geometric insights, we establish a one-to-one correspondence between maximal LDP channels under the Blackwell order and a finite-dimensional polytope, yielding an exact geometric characterization. This result renders the optimal PUT computationally tractable via vertex enumeration or linear programming. Furthermore, when the underlying problem exhibits symmetries characterized by a transitive group action, we derive an exact analytic expression for the optimal PUT, leading to closed-form solutions without numerical optimization. Our framework applies broadly beyond risk minimization, encompassing the maximization of information-theoretic measures such as mutual information, $f$-divergences, and Fisher information over LDP channels. We demonstrate the efficacy of our theoretical framework by recovering or strengthening several known results, and deriving exact analytic expressions for the optimal PUTs in specific tasks that were previously unaddressed.

Australia's New Payments Platform (NPP) processes 5.2 million real-time transactions per day under a 2,000 ms end-to-end SLA. With cryptographically relevant quantum computers projected for 2030-2035, financial institutions face a post-quantum cryptography (PQC) migration requirement. This paper presents a simulation study of the NIST FIPS 204 and 205 signature standards (ML-DSA, SLH-DSA/SPHINCS+) together with the NIST-selected Falcon algorithm (the forthcoming FIPS 206) in Australian payment infrastructure. Using signing latency measured with liboqs 0.14.0, we simulate 1,000 seasonally mixed days across NPP, RITS and SWIFT routes. All ML-DSA and Falcon variants achieve 100% SLA compliance, with a worst-case NPP p99 overhead of 1.57 ms (0.079% of the SLA budget). We introduce the Crypto Dilution Index (CDI), showing that PQC signing accounts for under 4% of end-to-end latency for every non-SPHINCS+ algorithm, so algorithm choice is not latency-driven. SPHINCS+ is the exception: at NPP volumes its signing time saturates a standard two-unit HSM pool, giving 0% NPP compliance, though it remains viable on the lower-rate RITS and SWIFT routes. A message-format analysis finds that SWIFT imposes no signature-size constraint. SWIFT authentication is applied out-of-band at the SWIFTNet messaging/transport layer, so neither the signature nor the verifying key travels in the Block-4 message text; every NIST PQC signature is size-compatible with SWIFT. Under a conservative harvest-now-decrypt-later scenario with a 2030 quantum arrival, roughly 9.56 billion NPP records generated between 2026 and 2029 would be retroactively exposed. We recommend ML-DSA-65 with ML-KEM-768 as the default across NPP and SWIFT, with Falcon-512 where genuinely in-band, size-constrained carriage is required.

Large language models (LLMs) face diverse privacy threats, yet existing studies typically evaluate individual attacks in isolation, offering limited guidance on which system factors drive privacy vulnerabilities. This paper presents a unified threat model and a systematic ablation study of four representative attacks: Membership Inference (MIA), Attribute Inference (AIA), Data Extraction (DEA), and Backdoor Attacks (BA). Our findings reveal attack-specific sensitivities: membership inference depends strongly on dataset characteristics and attack strategy, attribute inference is influenced by model and training characteristics, data extraction increases with model scale and training-data repetition, and backdoor attacks achieve consistently high success rates. Based on these findings, we derive practical recommendations for privacy-aware LLM deployment, including prioritizing membership leakage assessments in retrieval-augmented systems, minimizing repeated personal information in training data, monitoring attribute inference risks, and safeguarding fine-tuning pipelines against backdoors.

LLM agents convert model outputs into consequential actions, including communications, code changes, and financial transactions. Developers often trust evidence such as test results and execution logs. We identify a response path integrity gap in Bring Your Own Key configurations used by roughly 88 percent of mainstream agents. Because traffic passes through a user-authorized relay, the relay can modify plaintext LLM responses after alignment but before execution without breaking encryption. A minimal attack rewrites one execution bearing field and regenerates the remaining response using the user key while preserving the model style. Experiments reveal false green verification, where malicious code modifications pass public tests while silently defeating security checks. On APPS, 99.7 percent of publicly passing solutions retained downgraded behavior without developer-visible warnings. Tests on SWE bench, AgentDojo, and ASB across five frontier models show that single-field rewriting can redirect agents while preserving apparent task completion. We propose sign-c, a server-side scheme that authenticates execution bearing fields and outgoing queries. A local shim verifies them before action, while encryption protects confidentiality. The defense rejected all tampered responses with zero false rejections and only 0.0167 percent latency overhead.

When source code or the original toolchain is unavailable, patching binaries is difficult because it requires editing low-level assembly code directly. As an alternative, one can decompile the binary, apply the patch at the source level, and then recompile the modified code. However, as this paper demonstrates, this workflow is hindered by pervasive syntactic and semantic inaccuracies in the output of modern decompilers, many of which prior work has overlooked. To address these challenges, we present SCRIBE, a patching framework that handles syntactic and semantic issues in decompiled code, improving both recompilation success and correctness. SCRIBE's novel "binary-aware" recompilation approach repairs semantic inaccuracies in decompiler output by leveraging information extracted directly from the original binary. In our evaluation, SCRIBE resolved approximately 81% of previously incorrect functions produced by the Hex-Rays decompiler, demonstrating the effectiveness of its approach. Moreover, we show that, using SCRIBE, it is possible to patch 13 of 14 real-world CVEs without access to the original source code and without performing any manual binary editing. To further validate our findings, we conducted a user study with 18 participants. Using SCRIBE, participants achieved 100% patching success, compared to 3.7% without it. Finally, we asked three large language models to generate source-level patches via SCRIBE; all three achieved 100% success when using the framework, demonstrating its potential to enable fully automated patching. Overall, these results indicate that SCRIBE makes source-level patching of binaries accessible and reliable, even without access to the original source.

Federated learning (FL) is vulnerable to poisoning attacks, where malicious clients upload manipulated updates to degrade the performance of the global model. Although detection methods can identify and remove malicious clients, the model remains affected. Retraining from scratch is effective but costly, and existing unlearning methods remain unsatisfactory in both effectiveness and efficiency. We propose Federated Adversarial Unlearning (FAUN), a lightweight framework that retains only a short window of malicious clients' updates and employs adversarial optimization on a proxy dataset to derive updates that eliminate malicious directions. Applying these updates for a few unlearning rounds, followed by benign fine-tuning, enables fast removal of malicious effects and stable recovery. Experiments on three canonical datasets show that FAUN achieves recovery comparable to retraining while requiring far fewer rounds and reduces attack success rates to near zero, confirming FAUN successfully eliminates the contributions of unlearned clients.

The nonuniform and growing impact of adversarial noise across the layers of deep neural networks has been used in the literature, without a formal mathematical justification, to detect adversarial inputs and improve robustness. In this work, we study this phenomenon in detail and present a formal adversarial noise amplification theorem. We specify a set of sufficient conditions under which the adversarial noise amplification is mathematically guaranteed. Based on theoretical observations, we propose a novel training methodology with a custom spectral loss function and a specific architectural design to enhance the amplification signal for detecting adversarial data. Finally, we introduce a new, lightweight detection mechanism that leverages the enhanced amplification signal and operates entirely at inference time. To validate our approach, we demonstrate the detector's efficacy against both state-of-the-art attacks and a purpose-built adaptive attack, confirming that enhanced amplification can serve as a robust and reliable signal for adversarial defense.

ECDSA signatures form the bedrock of blockchain transaction authentication, yet their security critically depends on proper nonce generation. We uncover a critical vulnerability in the Polygon MEV ecosystem: systematic nonce reuse that enables complete private key recovery. Analyzing on-chain data reveals that searchers, driven by the need for sub-second response times in sealed-bid auctions, employ predictable nonce patterns. These patterns create linear relationships between signatures, allowing passive attackers to recover private keys using elementary algebra. We provide a compact linear-system formulation for such attacks, including the dangerous case of cross-wallet nonce collisions, and present concrete evidence of exploitable patterns on Polygon. Our findings demonstrate how protocol-induced latency pressures can lead to catastrophic cryptographic failures in production blockchain systems, where a single implementation error compromises multiple accounts simultaneously.

AI-based sanitization can remove concepts from images and text, but privacy evaluation remains largely ad hoc. We propose contrastive privacy, a formal definition that yields a quantitative test with a semantic interpretation. Under formal assumptions, we derive a conditional sufficiency result for a class of sanitized renderings (i.e., media files). We operationalize the definition using imperfect semantic-distance models such as CLIP. The test compares sanitized renderings under audit with both the original and sanitized versions of reference renderings known to contain privacy-relevant properties; if the rendering under audit is semantically closer to the unsanitized reference, then the former might leak private information even after sanitization. Importantly, the test is able to conditionally audit an abstract privacy target without enumerating its constituent properties or requiring per-item ground-truth labels; results remain relative to the chosen models and data. We evaluate 34 image-sanitization configurations, 15 social-media text models and one entity recognizer, and four off-the-shelf PII sanitizers on Enron emails. The tests detect residual semantic associations in every setting, including all four PII tools even when they sanitize every candidate their detectors return. Two further studies examine sensitivity to the semantic-distance model. For a synthetic identity, fine-tuning reveals target associations missed by the base model; after broader sanitization, the adapted test detects none. Across 94 matched image collections sanitized to conceal Leonardo DiCaprio's identity, three models yield broadly correlated assessments but sometimes disagree on which sanitizations appear most private. These findings support using multiple models and show how contrastive privacy can reveal retained identifiers and identity-revealing context across sanitization methods.

Network traffic anomaly detection represents a critical cybersecurity task, yet widespread encryption makes this task increasingly challenging. In response, image-based methods that model traffic as visual patterns have emerged as the dominant approach. However, this work pioneers the identification of a pervasive “full-frequency” characteristic and an associated limitation termed “spectral mismatch” within this paradigm. Specifically, while encrypted traffic exhibits prominent high-frequency components, mainstream reconstruction methods demonstrate an inherent bias toward learning low-frequency information. This fundamental mismatch results in incomplete representations that consequently degrade anomaly detection performance. To address this challenge, we propose FreeUp, a novel frequency-decoupled framework designed explicitly for encrypted traffic analysis. FreeUp decomposes traffic data into distinct low- and high-frequency bands, processing them through separate, dedicated branches along with a customized training strategy that ensures stable and independent frequency-specific learning. Furthermore, recognizing that simple reconstruction error proves inadequate for evaluating dual-branch architectures, we introduce an uncertainty-inspired fusion scoring mechanism. This mechanism quantifies the reconstruction uncertainty of the frequency-specific branches and dynamically integrates their outputs, yielding a more comprehensive and reliable anomaly score. Extensive experiments across multiple benchmarks demonstrate that FreeUp consistently outperforms state-of-the-art baselines. The code is available at https://github.com/ikun0124/FreeUp.