arXiv subjects

Olivera Kotevska

Publications and source records attributed to Olivera Kotevska.

At least 19 recordsLinked to original sources

arXiv2026

Forgetting Only What Matters: Layer-Selective Unlearning toward Robust LLMs

Large Language Models (LLMs) can memorize and reproduce sensitive, copyrighted, or otherwise undesirable training content, creating privacy, safety, and regulatory concerns. Machine unlearning offers a practical alternative to full retraining, but many existing methods apply broad or fixed parameter updates that can degrade utility and remain brittle under deployment changes such as post-training quantization, where forgotten knowledge may partially re-emerge. We propose Forgetting Only What Matters via Unlearning Layers (FOM-UL), a layer-level unlearning framework that selects transformer layers using a forget-to-retain significance score. This score identifies layers with high influence on the forget set and low sensitivity to the retain set, allowing FOM-UL to concentrate updates where they are most effective while leaving most of the model unchanged. This targeted update strategy improves the forgetting-utility trade-off and provides an empirical path toward quantization-resilient unlearning by reducing the chance that small, diffuse updates are erased by low-bit rounding. Across TOFU, KnowUnDo, and MUSE-style evaluations, FOM-UL reduces residual memorization compared with strong GA, NPO, KLD, SURE, ReLearn, and LUNAR-based baselines while preserving retain-set utility close to the vanilla model. Under 8-bit and 4-bit post-training quantization, FOM-UL maintains stronger memorization suppression and utility preservation than competing methods, and adversarial prompt evaluations show lower recovery of forgotten content. Overall, FOM-UL provides an efficient unlearning strategy that improves targeted forgetting, utility preservation, and deployment robustness without claiming formal guarantees of erasure.

cs.LG↗

arXiv2026

Extending the Horizon of Early Diagnosis: Lung Cancer Prediction with Vision Transformers

Lung cancer remains a leading cause of cancer-related mortality worldwide, and early diagnosis is critical for improving survival. However, early-stage malignancies can be subtle on chest X-rays, creating challenges for radiologists. This study evaluates Vision Transformers (ViTs) for predicting lung cancer one to two years before clinical diagnosis. We analyzed 259,361 chest X-rays from 91,020 imaging studies at the Jamaica Plains VA Hospital in Boston, MA. The dataset showed extreme class imbalance, approximately 1:150 cancer to non-cancer, which was addressed using hybrid under- and over-sampling and class-weighted loss optimization. Three ViT configurations were evaluated: a model trained from scratch, an ImageNet-pretrained model, and a Corona-pretrained model fine-tuned on the lung cancer dataset. Transfer learning improved performance, with pretrained models exceeding the scratch baseline by 6-10 percentage points in AUC and about 10-12 percent in balanced accuracy. ImageNet-pretrained models showed the most stable overall performance, while Corona-pretrained models achieved higher sensitivity in some settings but greater variability. Moderate resampling ratios, including 1:1 undersampling and 1.5:2 oversampling, provided favorable trade-offs between sensitivity, precision, and computational efficiency, reducing runtime by up to 70 percent without major performance loss. These findings demonstrate the potential of ViTs for early lung cancer risk prediction from routine chest X-rays. Although performance remains below clinical deployment thresholds, the results support further development of ViT-based triage systems to flag high-risk patients for earlier evaluation.

cs.CV↗

arXiv2026

Enforcing LLM Safety through DMD-based Classification of Prompt-Response Embedding Dynamics

Large Language Models (LLMs) are increasingly deployed in high-stakes applications, yet their tendency to generate toxic, harmful, or policy-violating content poses significant risks. Detecting these unsafe outputs efficiently in a black-box manner remains an open challenge. In this paper, we extend a recently proposed dynamical systems framework designed for hallucination detection to LLM safety classification. By projecting both prompts and responses into high-dimensional embedding spaces and fitting separate Koopman-based predictive models for safe and unsafe regimes, we classify new outputs using a new differential residual score that compares prediction errors of the safe and unsafe regimes. A key contribution is the incorporation of the prompt and response embedding dynamics, yielding fitted Koopman operators that capture crucial interaction patterns. We evaluate our black-box method across three safety benchmarks using three embedding models. Our results show that incorporating prompt embeddings yields consistent improvements, particularly for interaction-dependent violations when paired with causal decoders (e.g., in Llama-3), while response-only violations benefit more from dense semantic embedding representations. These findings opens the door for using dynamical systems to analyze AI systems rather than the dominant paradigm of using AI to model dynamical systems.

cs.AI↗

arXiv2026

Consilience: Conformally Calibrated Communication Control for Hidden-Profile Multi-Agent Reasoning

Multi-agent LLM systems can improve reasoning by pooling diverse perspectives, but their effectiveness depends on coordinating communication, particularly in hidden-profile settings where each agent holds only part of the evidence required for a correct decision. Existing protocols, including fixed schedules, round-robin exchange, and unstructured debate, provide no guarantee that a conversational action is appropriate. We propose Consilience, an inference-time orchestration framework that both steers and certifies multi-agent communication under distributed private information. At each turn, Consilience summarizes the discussion using a compact state capturing uncertainty, disagreement, evidence gain, redundancy, and premature consensus, then selects both a communication intervention (challenge, clarify, seek evidence, or route) and an appropriate speaker. Its central contribution is a round-wise conformal calibration procedure that provides a distribution-free, finite-sample guarantee: at each discussion round, conditional on reaching that round, the one-step regret of a controller's proposed action is bounded by a calibrated threshold with marginal probability at least 1 - alpha; an acceptance mechanism enforces the same guarantee for the executed action by replacing inadmissible proposals. On HiddenBench-style hidden-profile tasks spanning 12 open and closed weight language models, Consilience improves decision accuracy and communication efficiency over fixed and unstructured discussion protocols, sometimes surpassing a full-information baseline where every agent observes all evidence. These results demonstrate that certified adaptive communication control can be more valuable than increasing information availability, providing a practical mechanism for reliable multi-agent LLM coordination.

cs.AI↗

arXiv2026

IntraShuffler: A Privacy Preserving Framework for Heterogeneous DP Federated Learning

Heterogeneous Differential Privacy (HDP) in Federated Learning (FL) allows clients to select individual privacy budgets ($\varepsilon_i$) according to institutional policies and data sensitivity. In practice, many HDP-FL systems employ $\varepsilon$-aware server aggregation to improve model utility by re-weighting client updates according to their declared privacy budgets. However, gradient updates in FL retain structural patterns induced by non-independent and identically-distributed (non-IID) data, and these additional signals exposed by $\varepsilon$-aware aggregation create new opportunities for inference by an honest-but-curious server. In this work, we first show that a server equipped with gradient denoising and surrogate modeling can mount a \emph{Privacy Inference Attack} that infers distributional attributes of clients and links updates from the same client across training rounds, measured via surrogate inference accuracy and linkage success, under realistic knowledge constraints. The Shuffle-Model has been widely studied as a defense against such inference risks by anonymizing update sources, but it is fundamentally incompatible with HDP-FL $\varepsilon$-aware aggregation. To address this challenge, we propose \textbf{IntraShuffler}, a middleware defense framework designed for HDP-FL systems. IntraShuffler introduces a privacy-aware shuffling mechanism that groups clients into privacy-compatible buckets and performs parameter-level shuffling within each bucket to disrupt persistent gradient structure while preserving $\varepsilon$-aware aggregation. Experiments across four different datasets show that IntraShuffler reduces gradient recoverability by over 60% and decreases surrogate inference accuracy from 0.78 to 0.33 while maintaining comparable model utility across multiple FL aggregation rules.

cs.LG↗

arXiv2026

Vision Token Manipulation Attacks on Cloud-Edge Inference of Large Vision-Language Models

Cloud-edge Large Vision-Language Model (LVLM) inference enables efficient deployment by splitting computation between edge devices and cloud servers. In this process, intermediate vision tokens are transmitted from the edge to the cloud over a communication link, thereby exposing a new attack surface. We study vision token manipulation attack (VTM-Attack) under a black-box man-in-the-middle setting, where an adversary intercepts and manipulates a subset of transmitted vision tokens under a budget constraint. We propose four naïve attack strategies and an optimization-based token selection method. Experiments on 6 state-of-the-art LVLMs (3B-72B) across 4 benchmarks show that manipulating only 10\% of vision tokens can reduce accuracy by up to 88.31\%. These results reveal a critical vulnerability in cloud-edge LVLM inference.

cs.CR↗

arXiv2026

Closed-Loop Neural Activation Control in Vision-Language-Action Models

Vision-Language-Action (VLA) models can be steered at test time by intervening on semantically meaningful internal directions, but existing methods use a fixed steering coefficient, effectively operating in open loop. This is poorly suited to embodied control, where task state and concept error evolve over time, often causing overcorrection, oscillation, and reduced task success, especially for temporal behaviors such as speed and smoothness. We propose CTRL-STEER, a closed-loop framework that replaces static intervention strength with adaptive, time-varying control signals. The key idea is to decouple representation from regulation: rather than assuming temporal concepts are directly controlled by individual neurons, we steer along motion-aligned residual directions while a feedback controller adjusts intervention magnitude online. We instantiate this framework with both PID and reinforcement learning based controllers. Experiments with a fine-tuned OpenVLA policy on four LIBERO task suites show that CTRL-STEER achieves more stable concept regulation and a better steering-task success trade-off than fixed-coefficient baselines, without modifying or retraining the base model.

cs.AI↗

arXiv2026

SelfGrader: LLM Jailbreak Detection via Anchored Token-Level Logits

Large Language Models (LLMs) are powerful tools for answering user queries, yet they remain highly vulnerable to jailbreak attacks. Existing guardrail methods typically rely on internal features or textual responses to detect malicious queries, which either introduce substantial latency or suffer from randomness in text generation. To overcome these limitations, we propose SelfGrader, a lightweight guardrail method that formulates jailbreak detection as a numerical grading problem using anchored token-level logits. Specifically, SelfGrader evaluates the safety of a user query within a compact set of numerical tokens (NTs) (e.g., 0-9) and interprets their logit distribution as an internal safety signal. To align these signals with the target safety rubric, SelfGrader constructs Probably Approximately Correct-guided ICL anchor examples and introduces a dual-perspective scoring rule that considers both the maliciousness and benignness of the query, yielding a stable and interpretable score that reflects harmfulness and reduces the false positive rate simultaneously. Extensive experiments across diverse jailbreak benchmarks, adaptive attacks, benign prompt benchmarks, multiple LLMs, and state-of-the-art guardrail baselines demonstrate that SelfGrader achieves strong robustness with low false positive rates, memory overhead, and latency.

cs.CR↗

arXiv2026

Provably Communication-Efficient and Privacy-Preserving Federated Graph Neural Networks

Graph neural networks (GNNs) achieve strong performance on relational data, but real-world graphs are often distributed across organizations that cannot share raw data due to privacy and policy constraints. Existing federated GNN methods either ignore cross-client links, leading to degraded accuracy, or require frequent embedding exchanges, incurring substantial communication and privacy costs. We propose CE-FedGNN, a communication-efficient and privacy-preserving federated GNN framework for learning over such coupled graphs. Our approach avoids sharing raw data or per-round embeddings by infrequently exchanging aggregated node representations. To handle cross-client dependency and staleness, we introduce a moving-average estimator that continuously tracks node representations and enables their stable reuse across rounds. To provide formal privacy guarantees for the released representations, we adopt the metric differential privacy (metric-DP) framework, which measures privacy with respect to distances in the learned embedding space rather than worst-case input perturbations. This yields meaningful guarantees at noise levels where standard differential privacy becomes overly conservative. We establish convergence to a stationary point at a rate of $O(1/\sqrt{T})$ with $O(T^{3/4})$ communication complexity. In addition, we derive $(\varepsilon,δ)$-metric-DP guarantees via Rényi differential privacy composition under a public-cohort threat model. Experiments on synthetic interbank anti-money laundering benchmarks and citation networks demonstrate that CE-FedGNN achieves strong performance while significantly reducing communication and maintaining robustness under privacy-preserving noise.

cs.LG↗

arXiv2026

Majority Bit-Aware Watermarking For Large Language Models

The growing deployment of Large Language Models (LLMs) has raised concerns about their misuse in generating harmful or deceptive content. To address this issue, watermarking methods have been proposed to embed identifiable multi-bit messages into generated text for misuse tracing. However, existing methods often suffer from a fundamental trade-off between text quality and decoding accuracy. In particular, they have to restrict the size of the preferred token set (i.e., green list) during encoding to maintain a detectable watermark signal for decoding, which inevitably degrades generation quality. To improve this trade-off, we propose a novel message encoding paradigm called \textit{majority bit-aware encoding}, which relaxes the watermark signal strength from the green list size. This strategy allows for a strong watermark signal to be preserved in generated texts even when using a large green list. We introduce two instantiations of this paradigm: MajorMark and MajorMark$^{+}$, where the latter is specifically optimized for long messages. Extensive experiments on state-of-the-art LLMs demonstrate that our methods achieve higher decoding accuracy and superior text quality compared to prior baselines.

cs.CL↗

arXiv2026

XMark: Reliable Multi-Bit Watermarking for LLM-Generated Texts

Multi-bit watermarking has emerged as a promising solution for embedding imperceptible binary messages into Large Language Model (LLM)-generated text, enabling reliable attribution and tracing of malicious usage of LLMs. Despite recent progress, existing methods still face key limitations: some become computationally infeasible for large messages, while others suffer from a poor trade-off between text quality and decoding accuracy. Moreover, the decoding accuracy of existing methods drops significantly when the number of tokens in the generated text is limited, a condition that frequently arises in practical usage. To address these challenges, we propose \textsc{XMark}, a novel method for encoding and decoding binary messages in LLM-generated texts. The unique design of \textsc{XMark}'s encoder produces a less distorted logit distribution for watermarked token generation, preserving text quality, and also enables its tailored decoder to reliably recover the encoded message with limited tokens. Extensive experiments across diverse downstream tasks show that \textsc{XMark} significantly improves decoding accuracy while preserving the quality of watermarked text, outperforming prior methods. The code is at https://github.com/JiiahaoXU/XMark.

cs.CL↗

arXiv2026

Automated Membership Inference Attacks: Discovering MIA Signal Computations using LLM Agents

Membership inference attacks (MIAs), which enable adversaries to determine whether specific data points were part of a model's training dataset, have emerged as an important framework to understand, assess, and quantify the potential information leakage associated with machine learning systems. Designing effective MIAs is a challenging task that usually requires extensive manual exploration of model behaviors to identify potential vulnerabilities. In this paper, we introduce AutoMIA -- a novel framework that leverages large language model (LLM) agents to automate the design and implementation of new MIA signal computations. By utilizing LLM agents, we can systematically explore a vast space of potential attack strategies, enabling the discovery of novel strategies. Our experiments demonstrate AutoMIA can successfully discover new MIAs that are specifically tailored to user-configured target model and dataset, resulting in improvements of up to 0.18 in absolute AUC over existing MIAs. This work provides the first demonstration that LLM agents can serve as an effective and scalable paradigm for designing and implementing MIAs with SOTA performance, opening up new avenues for future exploration.

cs.CR↗

arXiv2026

Traceable Black-box Watermarks for Federated Learning

Due to the distributed nature of Federated Learning (FL) systems, each local client has access to the global model, which poses a critical risk of model leakage. Existing works have explored injecting watermarks into local models to enable intellectual property protection. However, these methods either focus on non-traceable watermarks or traceable but white-box watermarks. We identify a gap in the literature regarding the formal definition of traceable black-box watermarking and the formulation of the problem of injecting such watermarks into FL systems. In this work, we first formalize the problem of injecting traceable black-box watermarks into FL. Based on the problem, we propose a novel server-side watermarking method, $\mathbf{TraMark}$, which creates a traceable watermarked model for each client, enabling verification of model leakage in black-box settings. To achieve this, $\mathbf{TraMark}$ partitions the model parameter space into two distinct regions: the main task region and the watermarking region. Subsequently, a personalized global model is constructed for each client by aggregating only the main task region while preserving the watermarking region. Each model then learns a unique watermark exclusively within the watermarking region using a distinct watermark dataset before being sent back to the local client. Extensive results across various FL systems demonstrate that $\mathbf{TraMark}$ ensures the traceability of all watermarked models while preserving their main task performance. The code is available at https://github.com/JiiahaoXU/TraMark.

cs.CR↗

arXiv2026

Optimal Client Sampling in Federated Learning with Client-Level Heterogeneous Differential Privacy

Federated Learning with client-level differential privacy (DP) provides a promising framework for collaboratively training models while rigorously protecting clients' privacy. However, classic approaches like DP-FedAvg struggle when clients have heterogeneous privacy requirements, as they must uniformly enforce the strictest privacy level across all clients, leading to excessive DP noise and significant degradation in model utility. Existing methods to improve the model utility in such heterogeneous privacy settings often assume a trusted server and are largely heuristic, resulting in suboptimal performance and lacking strong theoretical foundations. In this work, we address these challenges under a practical attack model where both clients and the server are honest-but-curious. We propose GDPFed, which partitions clients into groups based on their privacy budgets and achieves client-level DP within each group to reduce the privacy budget waste and hence improve the model utility. Based on the privacy and convergence analysis of GDPFed, we find that the magnitude of DP noise depends on both model dimensionality and the per-group client sampling ratios. To further improve the performance of GDPFed, we introduce GDPFed$^+$, which integrates model sparsification to eliminate unnecessary noise and optimizes per-group client sampling ratios to minimize convergence error. Extensive empirical evaluations on multiple benchmark datasets demonstrate the effectiveness of GDPFed$^+$, showing substantial performance gains compared with state-of-the-art methods.

cs.CR↗

arXiv2025

Accuracy is Not Enough: Poisoning Interpretability in Federated Learning via Color Skew

As machine learning models are increasingly deployed in safety-critical domains, visual explanation techniques have become essential tools for supporting transparency. In this work, we reveal a new class of attacks that compromise model interpretability without affecting accuracy. Specifically, we show that small color perturbations applied by adversarial clients in a federated learning setting can shift a model's saliency maps away from semantically meaningful regions while keeping the prediction unchanged. The proposed saliency-aware attack framework, called Chromatic Perturbation Module, systematically crafts adversarial examples by altering the color contrast between foreground and background in a way that disrupts explanation fidelity. These perturbations accumulate across training rounds, poisoning the global model's internal feature attributions in a stealthy and persistent manner. Our findings challenge a common assumption in model auditing that correct predictions imply faithful explanations and demonstrate that interpretability itself can be an attack surface. We evaluate this vulnerability across multiple datasets and show that standard training pipelines are insufficient to detect or mitigate explanation degradation, especially in the federated learning setting, where subtle color perturbations are harder to discern. Our attack reduces peak activation overlap in Grad-CAM explanations by up to 35% while preserving classification accuracy above 96% on all evaluated datasets.

cs.CV↗

arXiv2025

HEMERA: A Human-Explainable Transformer Model for Estimating Lung Cancer Risk using GWAS Data

Lung cancer (LC) is the third most common cancer and the leading cause of cancer deaths in the US. Although smoking is the primary risk factor, the occurrence of LC in never-smokers and familial aggregation studies highlight a genetic component. Genetic biomarkers identified through genome-wide association studies (GWAS) are promising tools for assessing LC risk. We introduce HEMERA (Human-Explainable Transformer Model for Estimating Lung Cancer Risk using GWAS Data), a new framework that applies explainable transformer-based deep learning to GWAS data of single nucleotide polymorphisms (SNPs) for predicting LC risk. Unlike prior approaches, HEMERA directly processes raw genotype data without clinical covariates, introducing additive positional encodings, neural genotype embeddings, and refined variant filtering. A post hoc explainability module based on Layer-wise Integrated Gradients enables attribution of model predictions to specific SNPs, aligning strongly with known LC risk loci. Trained on data from 27,254 Million Veteran Program participants, HEMERA achieved >99% AUC (area under receiver characteristics) score. These findings support transparent, hypothesis-generating models for personalized LC risk assessment and early intervention.

cs.LG↗

arXiv2025

OmniFed: A Modular Framework for Configurable Federated Learning from Edge to HPC

Federated Learning (FL) is critical for edge and High Performance Computing (HPC) where data is not centralized and privacy is crucial. We present OmniFed, a modular framework designed around decoupling and clear separation of concerns for configuration, orchestration, communication, and training logic. Its architecture supports configuration-driven prototyping and code-level override-what-you-need customization. We also support different topologies, mixed communication protocols within a single deployment, and popular training algorithms. It also offers optional privacy mechanisms including Differential Privacy (DP), Homomorphic Encryption (HE), and Secure Aggregation (SA), as well as compression strategies. These capabilities are exposed through well-defined extension points, allowing users to customize topology and orchestration, learning logic, and privacy/compression plugins, all while preserving the integrity of the core system. We evaluate multiple models and algorithms to measure various performance metrics. By unifying topology configuration, mixed-protocol communication, and pluggable modules in one stack, OmniFed streamlines FL deployment across heterogeneous environments. Github repository is available at https://github.com/at-aaims/OmniFed.

cs.LG↗

arXiv2024

Improving Robustness of Spectrogram Classifiers with Neural Stochastic Differential Equations

Signal analysis and classification is fraught with high levels of noise and perturbation. Computer-vision-based deep learning models applied to spectrograms have proven useful in the field of signal classification and detection; however, these methods aren't designed to handle the low signal-to-noise ratios inherent within non-vision signal processing tasks. While they are powerful, they are currently not the method of choice in the inherently noisy and dynamic critical infrastructure domain, such as smart-grid sensing, anomaly detection, and non-intrusive load monitoring.

cs.LG↗