Searcharxiv⌕ Search

arXiv subjects

Marten van Dijk

Publications and source records attributed to Marten van Dijk.

At least 19 recordsLinked to original sources

PACZero: PAC-Private Fine-Tuning of Language Models via Sign Quantization

Differentially private (DP) zeroth-order (ZO) fine-tuning of large language models (LLMs) inherits DP's harsh utility-privacy trade-off. PAC Privacy, an instance-based notion, offers an alternative. It bounds the mutual information (MI) between the secret input and the output, and calibrates the noise to the stability of the output under the input distribution. We introduce PACZero, a family of ZO fine-tuning mechanisms under PAC Privacy. For stability, PACZero quantizes each update to a single bit, the sign of the ZO update aggregated over the secret subset of the training data. On many steps all candidate subsets agree on this sign, so the bit can be released at no cost in MI. On the remaining steps the bit passes through a noisy channel calibrated to a per-step MI budget. We study two channels, the standard additive Gaussian channel of PAC Privacy (PACZero-AGC) and a binary asymmetric channel with optimal bit flipping (PACZero-BAC), along with their zero-budget versions at $\mathrm{MI} \to 0$, PACZero-AGC$_0$ and PACZero-BAC$_0$. At zero MI, nothing released during training, including the fine-tuned model, carries any information about which secret subset was sampled, so no adversary inferring that subset can do better than its prior. In particular, the success rate of any membership inference attack (MIA) is bounded by 50% under a balanced prior. To our knowledge, this is the first such guarantee for LLM fine-tuning, and DP matches the MIA bound only at $\varepsilon = δ= 0$ and infinite noise. We evaluate on SST-2 and SQuAD with models from 1.3B to 8B parameters, under LoRA and full fine-tuning. Both channels retain high utility under tight MI budgets, and even PACZero-BAC$_0$, which guarantees zero MI, stays within 2 points of non-private ZO fine-tuning with the same number of steps. To our knowledge no DP method achieves usable utility matching this MIA-resistance level.

cs.LG↗

Closed-Form Noise Calibration Against Membership Inference for Random-Allocation DP-SGD

DP-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 $σ$, and with membership and non-membership equally likely a priori, the attack accuracy is at most $\frac12+\frac14\sqrt{(1+(e^{1/σ^2}-1)/M)^E-1}$. 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 $σ$ 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 $σ$ can certify a smaller $σ$. In training, the resulting $σ$ 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.

cs.LG↗

Trade-off Functions for DP-SGD with Subsampling based on Random Allocation: Tight Upper and Lower Bounds

Within 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 $σ\geq \sqrt{3/\ln M}$, where $σ$ 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 $\geq 1-a-δ$, that is, only $δ$ below the ideal random guessing diagonal $1-a$. For $δ= 1/100$ and $σ= 1$, roughly $M \approx 1.14\times 10^6$ rounds and $N \approx 1.14\times 10^7$ training samples suffice to achieve meaningful differential privacy. This is in contrast to recent negative results for the regime $σ\leq 1/\sqrt{2 \ln M}$ for which no significant DP guarantee can exist.

cs.LG↗

s-MDM: Generative Virtualization of Multi-Device Hardware Variations for Portable DL-SCA

Deep Learning-based Side-Channel Analysis (DL-SCA) frequently suffers from catastrophic performance degradation across unseen hardware due to printed circuit board routing differences, silicon process variations, and measurement noise shifts. This poster presents the Synthetic Multiple Device Model (s-MDM), a zero-target-trace generative framework designed to improve cross-device portability. s-MDM combines a structured cVAE generator, a Walsh-Hadamard leakage anchor, continuous style modulation, and decoupled leakage-style--domain critics to synthesize virtual source-device profiles offline. Benchmarked on 32-bit side-channel traces (AES_PTv2), s-MDM maps a precise operational boundary: while physical MDM remains superior on identical electrical clones (D4), s-MDM achieves consistently low key rank on the layout/acquisition-shifted Pinata target, where physical baselines are unstable or misaligned.

cs.CR↗

Fundamental Limitations of Favorable Privacy-Utility Guarantees for DP-SGD

Differentially Private Stochastic Gradient Descent (DP-SGD) is the dominant paradigm for private training, but its fundamental limitations under worst-case adversarial privacy definitions remain poorly understood. We analyze DP-SGD in the $f$-differential privacy framework, which characterizes privacy via hypothesis-testing trade-off curves, and study shuffled sampling over a single epoch with $M$ gradient updates. We derive an explicit suboptimal upper bound on the achievable trade-off curve. This result induces a geometric lower bound on the separation $κ$, which is the maximum distance between the mechanism's trade-off curve and the ideal random-guessing line. Because a large separation implies significant adversarial advantage, meaningful privacy requires small $κ$. However, we prove that enforcing a small separation imposes a strict lower bound on the Gaussian noise multiplier $σ$, which directly limits the achievable utility. In particular, under the standard worst-case adversarial model, shuffled DP-SGD must satisfy $$σ\ge \frac{1}{\sqrt{2\ln M}} \quad\text{or}\quad κ\ge\ \frac{1}{\sqrt{8}}\!\left(1-\frac{1}{\sqrt{4π\ln M}}\right),$$ thus cannot simultaneously achieve strong privacy and high utility. Although the noise threshold vanishes asymptotically as $M \to \infty$, the convergence is extremely slow. Even for practically relevant numbers of updates the required noise magnitude remains substantial. We further show that the same limitation extends to Poisson subsampling up to constant factors. Our experiments confirm that the noise levels implied by this bound lead to significant accuracy degradation at realistic training settings, thus showing a bottleneck in DP-SGD under standard worst-case adversarial assumptions.

cs.LG↗

Physical Time-Lock Puzzles

Traditional time-lock puzzles enforce delayed access to encrypted secrets by relying on inherently sequential computational steps. However, since software-based constructions impose no fundamental bound on the physical execution speed of individual steps, they remain vulnerable to hardware acceleration and improved implementations. In practice, this renders existing schemes ``step-lock'' rather than true ``time-lock'' puzzles, making long-term delay guarantees highly speculative against decades of unpredictable hardware advancement. To address this fundamental limitation, we introduce the Physical Time-Lock Puzzle (P-TLP), a new paradigm that anchors solving delay directly to the intrinsic, hardware-bounded evaluation latency of silicon hardware. By leveraging noisy Physical Unclonable Functions (PUFs) as non-parallelizable delay oracles, P-TLPs establish resistance against both parallel computation and algorithmic acceleration under the random oracle model. To tolerate PUF evaluation noise while maintaining a tightly concentrated and predictable solving-delay window, we propose a composite puzzle architecture that combines multiple independent basic puzzles. We formally prove the optimality of a sequential greedy solving strategy and derive tightly concentrated solving-delay windows using Hoeffding and Berry-Esseen bounds. We additionally prove early-solve hardness, establishing that adversaries with substantial classical computation budgets cannot bypass the physically-enforced PUF evaluation bottleneck. We validate our theoretical analysis with an FPGA prototype built around a configurable Ring Oscillator PUF, confirming that our analytical delay predictions tightly match empirical measurements. Our results demonstrate that P-TLPs can deliver highly predictable, long-term delay guarantees, enabling practical deployment in high-value applications such as digital legacy management.

cs.CR↗

Function Recovery Attacks in Gate-Hiding Garbled Circuits using SAT Solving

Semi-Private Function Evaluation (SPFE) enables joint computation while protecting both input data and the function itself. A practical instantiation is gate-hiding garbled circuits, which conceal gate functionalities while revealing circuit topology. Existing security definitions intentionally exclude leakage through topology, leaving its concrete impact on function privacy largely unexplored. We present a SAT-based function-recovery attack that reconstructs hidden gate operations from a circuit's public topology under two attacker knowledge models. Our approach combines topology-preserving simplification theorems with a decomposition of the recovery task into smaller SAT queries, thereby reducing the candidate gate-type assignment space and improving recovery performance. We evaluate the attack on ISCAS benchmarks, representative secure computation circuits, and fault-tolerant sensor fusion circuits under a 24-hour recovery budget. Compared to a baseline attack, the optimized version substantially reduces recovery time and, in some cases, completes recovery within the evaluation budget where the baseline does not. Our results show that revealing circuit topology can materially assist recovery of hidden gate functionality, identifying topology as a security-relevant leakage channel in gate-hiding garbled circuits.

cs.CR↗

On the Evidentiary Limits of Membership Inference for Copyright Auditing

As large language models (LLMs) are trained on increasingly opaque corpora, membership inference attacks (MIAs) have been proposed to audit whether copyrighted texts were used during training, despite growing concerns about their reliability under realistic conditions. We ask whether MIAs can serve as admissible evidence in adversarial copyright disputes where an accused model developer may obfuscate training data while preserving semantic content, and formalize this setting through a judge-prosecutor-accused communication protocol. To test robustness under this protocol, we introduce SAGE (Structure-Aware SAE-Guided Extraction), a paraphrasing framework guided by Sparse Autoencoders (SAEs) that rewrites training data to alter lexical structure while preserving semantic content and downstream utility. Our experiments show that state-of-the-art MIAs degrade when models are fine-tuned on SAGE-generated paraphrases, indicating that their signals are not robust to semantics-preserving transformations. While some leakage remains in certain fine-tuning regimes, these results suggest that MIAs are brittle in adversarial settings and insufficient, on their own, as a standalone mechanism for copyright auditing of LLMs.

cs.CR↗

DP-TLDM: Differentially Private Tabular Latent Diffusion Model

Synthetic data from generative models emerges as the privacy-preserving data sharing solution. Such a synthetic data set shall resemble the original data without revealing identifiable private information. Till date, the prior focus on limited types of tabular synthesizers and a small number of privacy attacks, particularly on Generative Adversarial Networks, and overlooks membership inference attacks and defense strategies, i.e., differential privacy. Motivated by the conundrum of keeping high data quality and low privacy risk of synthetic data tables, we propose DPTLDM, Differentially Private Tabular Latent Diffusion Model, which is composed of an autoencoder network to encode the tabular data and a latent diffusion model to synthesize the latent tables. Following the emerging f-DP framework, we apply DP-SGD to train the auto-encoder in combination with batch clipping and use the separation value as the privacy metric to better capture the privacy gain from DP algorithms. Our empirical evaluation demonstrates that DPTLDM is capable of achieving a meaningful theoretical privacy guarantee while also significantly enhancing the utility of synthetic data. Specifically, compared to other DP-protected tabular generative models, DPTLDM improves the synthetic quality by an average of 35% in data resemblance, 15% in the utility for downstream tasks, and 50% in data discriminability, all while preserving a comparable level of privacy risk.

cs.LG↗

Breaking XOR Arbiter PUFs with Chosen Challenge Attack

The XOR Arbiter PUF was introduced as a strong PUF in 2007 and was broken in 2015 by a Machine Learning (ML) attack, which allows the underlying Arbiter PUFs to be modeled individually by exploiting reliability information of the measured responses. To mitigate the reliability-based attacks, state-of-the-art understanding shows that the reliability of individual Arbiter PUFs and the overall XOR Arbiter PUF can be boosted to an arbitrarily high level, thus rendering all known reliability-based ML attacks infeasible; alternatively, an access control interface around the XOR Arbiter PUF can prevent the same challenge-response pairs from being accessed repeatedly, thus eliminating the leakage of reliability information. We show that, \textbf{for the first time, a perfectly reliable XOR Arbiter PUF can be successfully attacked in a divide-and-conquer manner}, meaning each underlying Arbiter PUF in an XOR Arbiter PUF can be attacked individually. This allows us to attack large XOR Arbiter PUFs efficiently, even without reliability information or any side-channel information. Our key insight is that, instead of reliability information, the responses of highly correlated challenges also reveal how close the responses are to the response decision boundary. This leads to a \textit{chosen challenge attack} on XOR Arbiter PUFs by carefully choosing correlated challenges to measure and aggregate the collected information. We validate our attack by using PUF simulation, as well as an XOR Arbiter PUF implemented on FPGA. We also demonstrate that our chosen challenge methodology is compatible with the state-of-the-art combined gradient-based multi-objective optimization attack. Finally, we discuss an effective countermeasure that can prevent our attack but with a relatively large area overhead compared to the PUF itself.

cs.CR↗

Beyond Anonymization: Object Scrubbing for Privacy-Preserving 2D and 3D Vision Tasks

We introduce ROAR (Robust Object Removal and Re-annotation), a scalable framework for privacy-preserving dataset obfuscation that eliminates sensitive objects instead of modifying them. Our method integrates instance segmentation with generative inpainting to remove identifiable entities while preserving scene integrity. Extensive evaluations on 2D COCO-based object detection show that ROAR achieves 87.5% of the baseline detection average precision (AP), whereas image dropping achieves only 74.2% of the baseline AP, highlighting the advantage of scrubbing in preserving dataset utility. The degradation is even more severe for small objects due to occlusion and loss of fine-grained details. Furthermore, 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. Our findings establish object removal as an effective privacy framework, achieving strong privacy guarantees with minimal performance trade-offs. The results highlight key challenges in generative inpainting, occlusion-robust segmentation, and task-specific scrubbing, setting the foundation for future advancements in privacy-preserving vision systems.

cs.CV↗

Proactive DP: A Multple Target Optimization Framework for DP-SGD

We introduce a multiple target optimization framework for DP-SGD referred to as pro-active DP. In contrast to traditional DP accountants, which are used to track the expenditure of privacy budgets, the pro-active DP scheme allows one to a-priori select parameters of DP-SGD based on a fixed privacy budget (in terms of $ε$ and $δ$) in such a way to optimize the anticipated utility (test accuracy) the most. To achieve this objective, we first propose significant improvements to the moment account method, presenting a closed-form $(ε,δ)$-DP guarantee that connects all parameters in the DP-SGD setup. We show that DP-SGD is $(ε<0.5,δ=1/N)$-DP if $σ=\sqrt{2(ε+\ln(1/δ))/ε}$ with $T$ at least $\approx 2k^2/ε$ and $(2/e)^2k^2-1/2\geq \ln(N)$, where $T$ is the total number of rounds, and $K=kN$ is the total number of gradient computations where $k$ measures $K$ in number of epochs of size $N$ of the local data set. We prove that our expression is close to tight in that if $T$ is more than a constant factor $\approx 4$ smaller than the lower bound $\approx 2k^2/ε$, then the $(ε,δ)$-DP guarantee is violated. The above DP guarantee can be enhanced in thatDP-SGD is $(ε, δ)$-DP if $σ= \sqrt{2(ε+\ln(1/δ))/ε}$ with $T$ at least $\approx 2k^2/ε$ together with two additional, less intuitive, conditions that allow larger $ε\geq 0.5$. Our DP theory allows us to create a utility graph and DP calculator. These tools link privacy and utility objectives and search for optimal experiment setups, efficiently taking into account both accuracy and privacy objectives, as well as implementation goals. We furnish a comprehensive implementation flow of our proactive DP, with rigorous experiments to showcase the proof-of-concept.

cs.LG↗

Recursive Augmented Fernet (RAF) Token: Alleviating the Pain of Stolen Tokens

A robust authentication and authorization mechanism is imperative in modular system development, where modularity and modular thinking are pivotal. Traditional systems often employ identity modules responsible for authentication and token issuance. Tokens, representing user credentials, offer advantages such as reduced reliance on passwords, limited lifespan, and scoped access. Despite these benefits, the "bearer token" problem persists, leaving systems vulnerable to abuse if tokens are compromised. We propose a token-based authentication mechanism addressing modular systems' critical bearer token problem. The proposed mechanism includes a novel RAF (Recursive Augmented Fernet) token, a blacklist component, and a policy enforcer component. RAF tokens are one-time-use tokens, like tickets. They carry commands, and the receiver of an RAF token can issue new tokens using the received RAF token. The blacklist component guarantees an RAF token can not be approved more than once, and the policy enforcer checks the compatibility of commands carried by an RAF token. We introduce two variations of RAF tokens: User-tied RAF, offering simplicity and compatibility, and Fully-tied RAF, providing enhanced security through service-specific secret keys. We thoroughly discuss the security guarantees, technical definitions, and construction of RAF tokens backed by game-based proofs. We demonstrate a proof of concept in the context of OpenStack, involving modifications to Keystone and creating an RAFT library. The experimental results reveal minimal overhead in typical scenarios, establishing the practicality and effectiveness of RAF. Our experiments show that the RAF mechanism beats the idea of using short-life Fernet tokens while providing much better security.

cs.CR↗

Generalizing DP-SGD with Shuffling and Batch Clipping

Classical differential private DP-SGD implements individual clipping with random subsampling, which forces a mini-batch SGD approach. We provide a general differential private algorithmic framework that goes beyond DP-SGD and allows any possible first order optimizers (e.g., classical SGD and momentum based SGD approaches) in combination with batch clipping, which clips an aggregate of computed gradients rather than summing clipped gradients (as is done in individual clipping). The framework also admits sampling techniques beyond random subsampling such as shuffling. Our DP analysis follows the $f$-DP approach and introduces a new proof technique which allows us to derive simple closed form expressions and to also analyse group privacy. In particular, for $E$ epochs work and groups of size $g$, we show a $\sqrt{g E}$ DP dependency for batch clipping with shuffling.

cs.LG↗

Considerations on the Theory of Training Models with Differential Privacy

In federated learning collaborative learning takes place by a set of clients who each want to remain in control of how their local training data is used, in particular, how can each client's local training data remain private? Differential privacy is one method to limit privacy leakage. We provide a general overview of its framework and provable properties, adopt the more recent hypothesis based definition called Gaussian DP or $f$-DP, and discuss Differentially Private Stochastic Gradient Descent (DP-SGD). We stay at a meta level and attempt intuitive explanations and insights \textit{in this book chapter}.

cs.LG↗

Game Theoretic Mixed Experts for Combinational Adversarial Machine Learning

Recent advances in adversarial machine learning have shown that defenses considered to be robust are actually susceptible to adversarial attacks which are specifically customized to target their weaknesses. These defenses include Barrage of Random Transforms (BaRT), Friendly Adversarial Training (FAT), Trash is Treasure (TiT) and ensemble models made up of Vision Transformers (ViTs), Big Transfer models and Spiking Neural Networks (SNNs). We first conduct a transferability analysis, to demonstrate the adversarial examples generated by customized attacks on one defense, are not often misclassified by another defense. This finding leads to two important questions. First, how can the low transferability between defenses be utilized in a game theoretic framework to improve the robustness? Second, how can an adversary within this framework develop effective multi-model attacks? In this paper, we provide a game-theoretic framework for ensemble adversarial attacks and defenses. Our framework is called Game theoretic Mixed Experts (GaME). It is designed to find the Mixed-Nash strategy for both a detector based and standard defender, when facing an attacker employing compositional adversarial attacks. We further propose three new attack algorithms, specifically designed to target defenses with randomized transformations, multi-model voting schemes, and adversarial detector architectures. These attacks serve to both strengthen defenses generated by the GaME framework and verify their robustness against unforeseen attacks. Overall, our framework and analyses advance the field of adversarial machine learning by yielding new insights into compositional attack and defense formulations.

cs.LG↗

Gradient Descent-Type Methods: Background and Simple Unified Convergence Analysis

In this book chapter, we briefly describe the main components that constitute the gradient descent method and its accelerated and stochastic variants. We aim at explaining these components from a mathematical point of view, including theoretical and practical aspects, but at an elementary level. We will focus on basic variants of the gradient descent method and then extend our view to recent variants, especially variance-reduced stochastic gradient schemes (SGD). Our approach relies on revealing the structures presented inside the problem and the assumptions imposed on the objective function. Our convergence analysis unifies several known results and relies on a general, but elementary recursive expression. We have illustrated this analysis on several common schemes.

math.OC↗

Finite-Sum Optimization: A New Perspective for Convergence to a Global Solution

Deep neural networks (DNNs) have shown great success in many machine learning tasks. Their training is challenging since the loss surface of the network architecture is generally non-convex, or even non-smooth. How and under what assumptions is guaranteed convergence to a \textit{global} minimum possible? We propose a reformulation of the minimization problem allowing for a new recursive algorithmic framework. By using bounded style assumptions, we prove convergence to an $\varepsilon$-(global) minimum using $\mathcal{\tilde{O}}(1/\varepsilon^3)$ gradient computations. Our theoretical foundation motivates further study, implementation, and optimization of the new algorithmic framework and further investigation of its non-standard bounded style assumptions. This new direction broadens our understanding of why and under what circumstances training of a DNN converges to a global minimum.

cs.LG↗