arXiv subjects

Yuhao He

Publications and source records attributed to Yuhao He.

10 recordsLinked to original sources

arXiv2026

Exploring and Bridging Knowledge Holes in Unlearned Multimodal Large Language Models

Machine unlearning offers a promising approach to remove unsafe content from Multimodal Large Language Models (MLLMs), yet ensuring the precision of unlearning remains a persistent challenge. One reason is that current MLLM unlearning evaluation paradigms suffer from a critical blind spot: they assess model utility through benchmarks whose representations are distant from the forget set, failing to capture knowledge holes---severe degradation on benign adjacent inputs. To probe knowledge holes in unlearned MLLMs, we construct a benchmark that captures unintended degradation on benign inputs sharing generic patterns with the forget set, and confirm through controlled experiments that they are a systematic consequence of commonly used approaches. Furthermore, to bridge this gap, we propose Selective Protection with Anchored Regularization, which protects generic patterns via anchored activation filtering while reinforcing them through entity-abstracted enhancement. Our experiments on SafeEraser demonstrate that SPAR recovers over 98% of vanilla response quality compared to below 50% for standard baselines---while achieving 0.00% attack success rate and competitive model utility. These results underscore the necessity of more fine-grained evaluation for trustworthy MLLM unlearning.

cs.AI↗

arXiv2026

Toward Fine-Grained Forgetting:Attribute Unlearning for Multimodal Large Language Models

Multimodal large language models (MLLMs) exhibit strong vision--language capabilities but may also memorize and disclose sensitive information. Machine unlearning seeks to remove designated knowledge without retraining from scratch while preserving general utility. Existing privacy-oriented benchmarks primarily adopt profile-level deletion, whereas practical requests are often finer grained: a model should forget a specified attribute while retaining non-sensitive information about the same identity. We therefore introduce attribute-level MLLM unlearning as a finer-grained task and construct a benchmark spanning long-text, numeric, and short-text targets, multiple forget ratios, and diverse question types. Our evaluation reveals that target and retained attributes share identity-specific and visual evidence, making selective forgetting susceptible to residual leakage or collateral degradation; accordingly, existing methods exhibit unstable forgetting--retention trade-offs in this setting. To address this challenge, we propose Causal Localization and Retain-Aware Projection (CLRP), a lightweight training-free framework. CLRP uses activation patching to identify the layer that causally mediates target-attribute disclosure, then applies a retain-aware projection that removes the target-attribute subspace while preserving same-identity evidence. Experiments across multiple widely used MLLMs with distinct architectures and parameter scales demonstrate the effectiveness of CLRP.

cs.AI↗

arXiv2026

Visual-Noise Guided In-Context Distillation for Multimodal Large Language Model Unlearning

Multimodal Large Language Models (MLLMs) have achieved remarkable progress on vision-language tasks, but they may also memorize and expose sensitive or restricted knowledge, raising concerns about privacy and broader safety risks. Machine Unlearning (MU) provides a promising way to remove targeted undesirable knowledge from trained models without retraining from scratch while preserving general model utility. Nevertheless, effective unlearning in MLLMs remains particularly challenging. Existing training-based methods often struggle to balance unlearning effectiveness and model utility. In contrast, training-free methods such as in-context unlearning preserve model utility by avoiding parameter updates, but they do not remove memorized knowledge at the parameter level and may remain vulnerable to reverse-engineering attacks. More importantly, in-context unlearning is insufficient in multimodal settings, where visual inputs can provide strong conditioning signals and induce undesirable outputs. To address these challenges, we propose Visual-Noise Guided In-Context Distillation (VGID), a distillation-based framework for MLLM unlearning. VGID dynamically constructs an unlearning-oriented teacher distribution from the frozen base model through dual-modal intervention that combines visual perturbation with textual in-context unlearning. The resulting intervention-induced distribution serves as a teacher signal for distillation, guiding the student model toward parameter-level unlearning without requiring external teacher models or explicit undesirable response annotations. Experimental results show that VGID achieves strong unlearning effectiveness while preserving competitive model utility, reducing forget set ROUGE-L by 0.371 with only a 0.055 drop in retain set ROUGE-L in a representative setting.

cs.CV↗

arXiv2025

ODIA: Oriented Distillation for Inline Acceleration of LLM-based Function Calling

Function Calling is a crucial technique that enables Large Language Models (LLMs) to interact with external systems through APIs. However, the high latency associated with LLM-based Function Calling significantly impacts user experience. This paper presents a novel approach called Oriented Distillation for Inline Acceleration (ODIA) that leverages online user interaction data to accelerate Function Calling. By automatically identifying "simple queries" from production traffic and distilling knowledge from larger models to smaller ones, our method reduces response latency by 45% (expected) and 78% (median) while maintaining accuracy. We demonstrate the effectiveness of our approach through real-world deployment in a music application, where the smaller model successfully handles 60% of traffic with negligible accuracy loss. Our method requires minimal human intervention and continuously improves through automated data collection and model updating, making it a practical solution for production environments.

cs.LG↗

arXiv2025

Structure Disruption: Subverting Malicious Diffusion-Based Inpainting via Self-Attention Query Perturbation

The rapid advancement of diffusion models has enhanced their image inpainting and editing capabilities but also introduced significant societal risks. Adversaries can exploit user images from social media to generate misleading or harmful content. While adversarial perturbations can disrupt inpainting, global perturbation-based methods fail in mask-guided editing tasks due to spatial constraints. To address these challenges, we propose Structure Disruption Attack (SDA), a powerful protection framework for safeguarding sensitive image regions against inpainting-based editing. Building upon the contour-focused nature of self-attention mechanisms of diffusion models, SDA optimizes perturbations by disrupting queries in self-attention during the initial denoising step to destroy the contour generation process. This targeted interference directly disrupts the structural generation capability of diffusion models, effectively preventing them from producing coherent images. We validate our motivation through visualization techniques and extensive experiments on public datasets, demonstrating that SDA achieves state-of-the-art (SOTA) protection performance while maintaining strong robustness.

cs.CV↗

arXiv2024

Deferred Poisoning: Making the Model More Vulnerable via Hessian Singularization

Recent studies have shown that deep learning models are very vulnerable to poisoning attacks. Many defense methods have been proposed to address this issue. However, traditional poisoning attacks are not as threatening as commonly believed. This is because they often cause differences in how the model performs on the training set compared to the validation set. Such inconsistency can alert defenders that their data has been poisoned, allowing them to take the necessary defensive actions. In this paper, we introduce a more threatening type of poisoning attack called the Deferred Poisoning Attack. This new attack allows the model to function normally during the training and validation phases but makes it very sensitive to evasion attacks or even natural noise. We achieve this by ensuring the poisoned model's loss function has a similar value as a normally trained model at each input sample but with a large local curvature. A similar model loss ensures that there is no obvious inconsistency between the training and validation accuracy, demonstrating high stealthiness. On the other hand, the large curvature implies that a small perturbation may cause a significant increase in model loss, leading to substantial performance degradation, which reflects a worse robustness. We fulfill this purpose by making the model have singular Hessian information at the optimal point via our proposed Singularization Regularization term. We have conducted both theoretical and empirical analyses of the proposed method and validated its effectiveness through experiments on image classification tasks. Furthermore, we have confirmed the hazards of this form of poisoning attack under more general scenarios using natural noise, offering a new perspective for research in the field of security.

cs.LG↗

arXiv2014

Characterization of superconducting nanowire single-photon detector with artificial constrictions

Statistical studies on the performance of different superconducting nanowire single-photon detectors (SNSPDs) on one chip suggested that random constrictions existed in the nanowire that were barely registered by scanning electron microscopy. With the aid of advanced e-beam lithography, artificial geometric constrictions were fabricated on SNSPDs as well as single nanowires. In this way, we studied the influence of artificial constrictions on SNSPDs in a straight forward manner. By introducing artificial constrictions with different wire widths in single nanowires, we concluded that the dark counts of SNSPDs originate from a single constriction. Further introducing artificial constrictions in SNSPDs, we studied the relationship between detection efficiency and kinetic inductance and the bias current, confirming the hypothesis that constrictions exist in SNSPDs.

cond-mat.supr-con↗

arXiv2013

Controlled entanglement routing between two virtual pathways

We demonstrate controlled entanglement routing between bunching and antibunching path-entangled two-photon states in an unbalanced Mach-Zehnder interferometer (UMZI), in which the routing process is controlled by the relative phase difference in the UMZI. Regarding bunching and antibunching path-entangled two-photon states as two virtual ports, we can consider the UMZI as a controlled entanglement router, which bases on the coherent manipulation of entanglement. Half of the entanglement within the input two-photon state is coherently routed between the two virtual ports, while the other is lost due to the time distinguishability introduced by the UMZI. Pure bunching or antibunching path entangled two-photon states are obtained based on this controlled entanglement router. The results show that we can employ the UMZI as general entanglement router for practical quantum information application.

quant-ph↗

arXiv2013

Jitter analysis of a superconducting nanowire single photon detector

Jitter is one of the key parameters for a superconducting nanowire single photon detector (SNSPD). Using an optimized time-correlated single photon counting system for jitter measurement, we extensively studied the dependence of system jitter on the bias current and working temperature. The signal-to-noise ratio of the single-photon-response pulse was proven to be an important factor in system jitter. The final system jitter was reduced to 18 ps by using a high-critical-current SNSPD, which showed an intrinsic SNSPD jitter of 15 ps. A laser ranging experiment using a 15-ps SNSPD achieved a record depth resolution of 3 mm at a wavelength of 1550 nm.

cond-mat.supr-con↗

arXiv2013

Time-of-flight laser ranging and imaging at 1550 nm using low-jitter superconducting nanowire single-photon detection system

We developed a time-correlated single-photon counting (TCSPC) system based on the low-jitter superconducting nanowire single-photon detection (SNSPD) technology. The causes of jitters in the TCSPC system were analyzed. Owing to the low jitter of the SNSPD technology, a system jitter of 26.8 ps full-width at half-maximum was achieved after optimizing the system. We demonstrated time-of-flight laser ranging at 1550 nm wavelength at a stand-off distance of 115 m, based on this TCSPC system. A depth resolution of 4 mm was achieved directly by locating the centroids of each of the two return signals. Laser imaging was also performed using the TCSPC system. This low-jitter TCSPC system using the SNSPD technology presents great potential in long-range measurements and imaging applications for low-energy-level and eye-safe laser systems

physics.ins-det↗