publications
publications by categories in reversed chronological order. generated by jekyll-scholar.
2026
- Preprint
PACZero: PAC-Private Fine-Tuning of Language Models via Sign QuantizationMurat Bilgehan Ertan, Xiaochen Zhu, Phuong Ha Nguyen, Marten van Dijk, and Srinivas DevadasAccepted at NeurIPS’26 (poster), 2026We introduce PACZero, a suite of PAC-private zeroth-order mechanisms for fine-tuning large language models that delivers usable utility at I(S*; Y_1:T) = 0, providing membership-inference resistance comparable to differential privacy only at extreme noise levels. The key idea is to sign-quantize subset-aggregated zeroth-order gradients to induce frequent unanimity—instances where candidate subsets agree on update directions—which incurs zero conditional mutual information cost. Two variants balance the privacy-utility spectrum: PACZero-MI (budgeted mutual information) and PACZero-ZPL (zero mutual information through randomized disagreement handling). Experiments on SST-2 and SQuAD with OPT models demonstrate competitive performance in high-privacy regimes where prior methods struggle.
@article{ertan2026paczero, author = {Ertan, Murat Bilgehan and Zhu, Xiaochen and Nguyen, Phuong Ha and {van Dijk}, Marten and Devadas, Srinivas}, title = {{PACZero}: {PAC}-Private Fine-Tuning of Language Models via Sign Quantization}, journal = {Accepted at NeurIPS'26 (poster)}, volume = {abs/2605.06505}, year = {2026}, doi = {10.48550/ARXIV.2605.06505}, } - Preprint
Fundamental Limitations of Favorable Privacy-Utility Guarantees for DP-SGDMurat Bilgehan Ertan and Marten van DijkAccepted at ACM CCS 2026, 2026This work establishes fundamental limitations of favorable privacy-utility guarantees for DP-SGD, providing theoretical analysis of the inherent trade-offs in differentially private stochastic gradient descent.
@article{ertan2026fundamental, author = {Ertan, Murat Bilgehan and {van Dijk}, Marten}, title = {Fundamental Limitations of Favorable Privacy-Utility Guarantees for DP-SGD}, journal = {Accepted at ACM CCS 2026}, volume = {abs/2601.10237}, year = {2026}, doi = {10.48550/ARXIV.2601.10237}, } - Preprint
Closed-Form Noise Calibration Against Membership Inference for Random-Allocation DP-SGDMurat Bilgehan Ertan and Marten van DijkCoRR, 2026DP-SGD protects training data by adding Gaussian noise to clipped gradients. The amount of noise is usually chosen by running a numerical privacy accountant inside a search. We study DP-SGD with random allocation, where each epoch uses every record once, at a randomly chosen step. For this setting we give a one-line formula that bounds the accuracy of every membership inference attack (MIA) on the trained model. With M steps per epoch, E epochs and noise multiplier sigma, and with membership and non-membership equally likely a priori, the attack accuracy is at most 1/2 + (1/4) sqrt((1 + (exp(1/sigma^2) - 1)/M)^E - 1). The formula comes from the chi-square divergence between a Gaussian distribution and a Gaussian mixture that dominates random allocation. It is interpretable and gives sigma in about a microsecond. Where applicable, our formula needs at most about half the noise of the state-of-the-art closed-form bound. To measure how close the bound is, we also derive an exact expression for the attack accuracy of these two distributions and evaluate it numerically. Calibrating to this exact expression requires 13.0% to 20.2% less noise than the formula in our main experiments, and since it is exact, no accountant that knows only M, E and sigma can certify a smaller sigma. In training, the resulting sigma outperforms the formula and matches a published accountant in test accuracy. It is found in seconds and certified in minutes, whereas every search we ran with that accountant took longer or returned at least 0.62% more noise. We show that MIAs on the trained models stay below the bound.
@article{ertan2026closedform, author = {Ertan, Murat Bilgehan and {van Dijk}, Marten}, title = {{Closed-Form} Noise Calibration Against Membership Inference for {Random-Allocation} {DP-SGD}}, journal = {CoRR}, volume = {abs/2610.09651}, year = {2026}, doi = {10.48550/ARXIV.2610.09651}, } - IEEE S&P
Beyond Stop Signs: Why Evasion Attacks Matter Even MoreKaleel Mahmood, Murat Bilgehan Ertan, and Marten van DijkIEEE Security & Privacy, 2026@article{mahmood2026beyond, author = {Mahmood, Kaleel and Ertan, Murat Bilgehan and {van Dijk}, Marten}, title = {Beyond Stop Signs: Why Evasion Attacks Matter Even More}, journal = {IEEE Security \& Privacy}, volume = {24}, number = {3}, pages = {65--70}, year = {2026}, publisher = {IEEE}, } - Preprint
Trade-off Functions for DP-SGD with Subsampling based on Random Allocation: Tight Upper and Lower BoundsMarten van Dijk and Murat Bilgehan ErtanCoRR, 2026Within the f-DP framework, we derive a tight analysis of the trade-off function for Differentially Private Stochastic Gradient Descent (DP-SGD) with subsampling based on random allocation, in which each sample is independently assigned to exactly one of M minibatches per epoch, each minibatch corresponding to one of the M SGD rounds within a single epoch. Our analysis holds under an explicit validity condition, whose hypotheses together force sigma >= sqrt(3/ln M), where sigma is the DP noise multiplier. Unlike f-DP analyses for Poisson subsampling, which yield non-closed implicit formulas that can be machine computed but are non-transparent, random allocation admits a tight analysis yielding transparent and interpretable closed-form bounds. For a single epoch, our concrete bounds, derived via the Berry-Esseen theorem, are tight up to constant factors. We demonstrate worked parameter settings for a single epoch (E=1) with a corresponding trade-off function >= 1 - a - delta, that is, only delta below the ideal random guessing diagonal 1 - a. For delta = 1/100 and sigma = 1, roughly M 1.14 x 10^6 rounds and N 1.14 x 10^7 training samples suffice to achieve meaningful differential privacy. This is in contrast to recent negative results for the regime sigma <= 1/sqrt(2 ln M) for which no significant DP guarantee can exist.
@article{vandijk2026tradeoff, author = {{van Dijk}, Marten and Ertan, Murat Bilgehan}, title = {Trade-off Functions for {DP-SGD} with Subsampling based on Random Allocation: Tight Upper and Lower Bounds}, journal = {CoRR}, volume = {abs/2605.06259}, year = {2026}, doi = {10.48550/ARXIV.2605.06259}, } - Preprint
On the Evidentiary Limits of Membership Inference for Copyright AuditingMurat Bilgehan Ertan, Emirhan Böge, Min Chen, Kaleel Mahmood, and Marten van DijkCoRR, 2026This work examines the evidentiary limits of membership inference attacks for copyright auditing, questioning their reliability as evidence of unauthorized training data usage.
@article{ertan2026evidentiary, author = {Ertan, Murat Bilgehan and B\"{o}ge, Emirhan and Chen, Min and Mahmood, Kaleel and {van Dijk}, Marten}, title = {On the Evidentiary Limits of Membership Inference for Copyright Auditing}, journal = {CoRR}, volume = {abs/2601.12937}, year = {2026}, doi = {10.48550/ARXIV.2601.12937}, } - ACM IWSPA
TOSSS: a CVE-based Software Security Benchmark for Large Language ModelsMarc Damie, Murat Bilgehan Ertan, Domenico Essoussi, Angela Makhanu, Gaëtan Peter, and Roos WensveenIn Proceedings of the 12th ACM International Workshop on Security and Privacy Analytics (IWSPA ’26), 2026TOSSS is a CVE-based software security benchmark designed to evaluate the capabilities of large language models in identifying and reasoning about software vulnerabilities.
@inproceedings{damie2026tosss, author = {Damie, Marc and Ertan, Murat Bilgehan and Essoussi, Domenico and Makhanu, Angela and Peter, Ga\"{e}tan and Wensveen, Roos}, title = {{TOSSS}: a {CVE}-based Software Security Benchmark for Large Language Models}, booktitle = {Proceedings of the 12th ACM International Workshop on Security and Privacy Analytics (IWSPA '26)}, pages = {111--117}, year = {2026}, publisher = {Association for Computing Machinery}, doi = {10.1145/3806007.3810968}, }
2025
- Master’s Thesis
Beyond Anonymization: Object Scrubbing for Privacy-Preserving 2D and 3D Vision TasksMurat Bilgehan Ertan, Ronak Sahu, Phuong Ha Nguyen, Kaleel Mahmood, and Marten van DijkCoRR, 2025We introduce ROAR (Robust Object Removal and Re-annotation), a scalable framework for privacy-preserving dataset obfuscation that removes sensitive objects instead of modifying them. Designed for practical deployment, our method integrates instance segmentation with generative inpainting to eliminate identifiable entities while preserving scene integrity. Extensive evaluations on 2D COCO-based object detection show that ROAR achieves 87.5% of baseline average precision (AP), whereas image dropping achieves only 74.2%, highlighting the advantage of scrubbing in preserving dataset utility. In NeRF-based 3D reconstruction, our method incurs a PSNR loss of at most 1.66 dB while maintaining SSIM and improving LPIPS, demonstrating superior perceptual quality.
@article{ertan2025beyond, author = {Ertan, Murat Bilgehan and Sahu, Ronak and Nguyen, Phuong Ha and Mahmood, Kaleel and {van Dijk}, Marten}, title = {Beyond Anonymization: Object Scrubbing for Privacy-Preserving 2D and 3D Vision Tasks}, journal = {CoRR}, volume = {abs/2504.16557}, year = {2025}, doi = {10.48550/ARXIV.2504.16557}, } - xAI
A Biologically Inspired Filter Significance Assessment Method for Model ExplanationEmirhan Böge, Yasemin Gunindi, Murat Bilgehan Ertan, Erchan Aptoula, Nihan Alp, and Huseyin OzkanIn Explainable Artificial Intelligence, 2025The interpretability of deep learning models remains a significant challenge, particularly in convolutional neural networks (CNNs) where understanding the contributions of individual filters is crucial for explainability. In this work, we propose a biologically inspired filter significance assessment method based on Steady-State Visually Evoked Potentials (SSVEPs), a well-established neuroscience principle. Our approach leverages frequency tagging techniques to quantify the importance of convolutional filters by analyzing their frequency-locked responses to periodic contrast modulations in input images. By blending SSVEP-based filter selection into Class Activation Mapping (CAM) frameworks such as Grad-CAM, Grad-CAM++, EigenCAM, and LayerCAM, we enhance model interpretability while reducing attribution noise.
@inproceedings{boge2025biologically, author = {B{\"o}ge, Emirhan and Gunindi, Yasemin and Ertan, Murat Bilgehan and Aptoula, Erchan and Alp, Nihan and Ozkan, Huseyin}, title = {A Biologically Inspired Filter Significance Assessment Method for Model Explanation}, booktitle = {Explainable Artificial Intelligence}, editor = {Guidotti, Riccardo and Schmid, Ute and Longo, Luca}, publisher = {Springer Nature Switzerland}, address = {Cham}, pages = {422--435}, isbn = {978-3-032-08324-1}, year = {2025}, }
2024
- ACM DTRAP
Unveiling Cyber Threat Actors: A Hybrid Deep Learning Approach for Behavior-based AttributionEmirhan Böge, Murat Bilgehan Ertan, Halit Alptekin, and Orçun ÇetinDigital Threats: Research and Practice, 2024In this paper, we leverage natural language processing and machine learning algorithms to profile threat actors based on their behavioral signatures to establish identification for soft attribution. Our unique dataset comprises various actors and the commands they have executed, with a significant proportion using the Cobalt Strike framework in August 2020-October 2022. We implemented a hybrid deep learning structure combining transformers and convolutional neural networks to benefit global and local contextual information within the sequence of commands, which provides a detailed view of the behavioral patterns of threat actors.
@article{boge2024unveiling, author = {B\"{o}ge, Emirhan and Ertan, Murat Bilgehan and Alptekin, Halit and \c{C}etin, Or\c{c}un}, title = {Unveiling Cyber Threat Actors: A Hybrid Deep Learning Approach for Behavior-based Attribution}, journal = {Digital Threats: Research and Practice}, year = {2024}, publisher = {Association for Computing Machinery}, address = {New York, NY, USA}, doi = {10.1145/3676284}, }