Post-training compression of Multimodal LLMs faces a fundamental geometric conflict: parameter subspaces optimized for text often suppress orthogonal visual features. We demonstrate that standard SVD fails to resolve this cross-modal mismatch, causing catastrophic visual degradation. To bridge this gap, we introduce Joint-Whitening SVD (JW-SVD), a dual-objective framework that aligns vision and language manifolds via a Joint Covariance basis, preserving features critical to both. Additionally, we propose Global Spectrum-Aware Truncation to dynamically transfer parameter budget from the redundant Vision Tower to the sensitive Backbone. Experiments on Qwen2.5-VL and Llama-3-Next confirm that JW-SVD demonstrates superior retention of both text and image capabilities. In addition, it resolves the modality trade-off: it recovers over 30% of perceptual performance lost by baselines while maintaining parity in textual reasoning, enabling robust multimodal performance even at extreme compression rates.

Edge deployment imposes strict latency, memory, and energy constraints that scale directly with network depth, yet the question of which architectural paradigm offers the best accuracy-efficiency tradeoff at ultra-low depth (four to six layers) remains open. We present the first systematic comparison of convolutional, pure transformer, and hybrid CNN-transformer models under fixed shallow depth budgets. We develop an analytical framework characterizing the representational cost of local convolutional versus global attention-based feature mixing as a function of depth, and validate it with experiments on ImageNet-1K across architectures spanning MobileNetV2, DeiT, Swin, MobileViT, and EfficientFormer. Beyond FLOPs and accuracy, we report real-world latency and energy measurements on an NVIDIA Jetson Nano and examine deployment feasibility on Cortex-M class microcontrollers. Our experiments reveal how accuracy, latency, and memory scale across all three architecture families as depth decreases, providing practitioners with direct guidance on which model class to choose for a given depth and hardware budget.

Quantization enables efficient deployment of large language models (LLMs) in resourceconstrained environments by significantly reducing memory and computation costs.While quantized LLMs often maintain performance on perplexity and zero-shot tasks, their impact on truthfulness-whether generating truthful or deceptive responses-remains largely unexplored.In this work, we introduce Truthful-nessEval, a comprehensive evaluation framework for assessing the truthfulness of quantized LLMs across three dimensions: (1) Truthfulness on Logical Reasoning; (2) Truthfulness on Common Sense; and (3) Truthfulness on Imitative Falsehoods.Using this framework, we examine mainstream quantization techniques (ranging from 4-bit to extreme 2-bit) across several open-source LLMs.Surprisingly, we find that while quantized models retain internally truthful representations, they are very susceptible to producing false outputs under misleading prompts.To probe this vulnerability, we test 15 rephrased variants of "honest", "neutral" and "deceptive" prompts and observe that "deceptive" prompts can override truth-consistent behavior, whereas "honest" and "neutral" prompts maintain stable outputs.Further, we reveal that quantized models "know" the truth internally yet still produce false outputs when guided by "deceptive" prompts via layer-wise probing.Our findings provide insights into future designs of trustworthy quantization-aware alignment.Codes and data are available here 1 .

Large language models (LLMs) tend to follow maliciously crafted instructions to generate deceptive responses, posing safety challenges.How deceptive instructions alter the internal representations of LLM compared to truthful ones remains poorly understood beyond output analysis.To bridge this gap, we investigate when and how these representations "flip", such as from truthful to deceptive, under deceptive versus truthful/neutral instructions.Analyzing the internal representations of Llama-3.1-8B-Instruct and Gemma-2-9B-Instruct on a factual verification task, we find the model's instructed True/False output is predictable via linear probes across all conditions based on the internal representation.Further, we use Sparse Autoencoders (SAEs) to show that the Deceptive instructions induce significant representational shifts compared to Truthful/Neutral representations (which are similar), concentrated in early-to-mid layers and detectable even on complex datasets.We also identify specific SAE features highly sensitive to deceptive instruction and use targeted visualizations to confirm distinct truthful/deceptive representational subspaces.Our findings expose featureand layer-level signatures of deception, offering new insights for detecting and mitigating instructed dishonesty in LLMs.The code is available at: https://github.com/ivyllll/truthful-representation-flip.

A large amount of high-dimensional and heterogeneous data appear in practical applications, which are often published to third parties for data analysis, recommendations, targeted advertising, and reliable predictions. However, publishing these data may disclose personal sensitive information, resulting in an increasing concern on privacy violations. Privacy-preserving data publishing has received considerable attention in recent years. Unfortunately, the differentially private publication of high dimensional data remains a challenging problem. In this paper, we propose a differentially private high-dimensional data publication mechanism (DP2-Pub) that runs in two phases: a Markov-blanket-based attribute clustering phase and an invariant post randomization (PRAM) phase. Specifically, splitting attributes into several low-dimensional clusters with high intra-cluster cohesion and low inter-cluster coupling helps obtain a reasonable allocation of privacy budget, while a double-perturbation mechanism satisfying local differential privacy facilitates an invariant PRAM to ensure no loss of statistical information and thus significantly preserves data utility. We also extend our DP2-Pub mechanism to the scenario with a semi-honest server which satisfies local differential privacy. We conduct extensive experiments on four real-world datasets and the experimental results demonstrate that our mechanism can significantly improve the data utility of the published data while satisfying differential privacy.

Conventional private data publication mechanisms aim to retain as much data utility as possible while ensuring sufficient privacy protection on sensitive data. Such data publication schemes implicitly assume that all data analysts and users have the same data access privilege levels. However, it is not applicable for the scenario that data users often have different levels of access to the same data, or different requirements of data utility. The multi-level privacy requirements for different authorization levels pose new challenges for private data publication. Traditional PPDP mechanisms only publish one perturbed and private data copy satisfying some privacy guarantee to provide relatively accurate analysis results. To find a good tradeoff between privacy preservation level and data utility itself is a hard problem, let alone achieving multi-level data utility on this basis. In this paper, we address this challenge in proposing a novel framework of data publication with compressive sensing supporting multi-level utility-privacy tradeoffs, which provides differential privacy. Specifically, we resort to compressive sensing (CS) method to project a n-dimensional vector representation of users’ data to a lower m-dimensional space, and then add deliberately designed noise to satisfy differential privacy. Then, we selectively obfuscate the measurement vector under compressive sensing by adding linearly encoded noise, and provide different data reconstruction algorithms for users with different authorization levels. Extensive experimental results demonstrate that ML-DPCS yields multi-level of data utility for specific users at different authorization levels.

Online social networks have gained tremendous popularity and have dramatically changed the way we communicate in recent years. However, the publishing of social network data raises more and more privacy concerns. To protect user privacy, social networking data are usually anonymized before being released. Nevertheless, existing anonymization techniques do not have sufficient protection effects. A large number of deanonymization attacks have arisen, and they mainly make use of either network topology or node attribute information to successfully reidentify anonymized users. In this article, we model a social network as a structure-attribute network (SAN) integrating the structural characteristics and the attribute information associated with social network users. A novel similarity measurement of social network nodes is proposed by considering the structural similarity and attribute similarity. A two-phase scheme is then designed to perform deanonymization by first dividing a social network (graph) into smaller subgraphs based on spectral graph partitioning and then applying the proposed deanonymization algorithm on each matched subgraph pair. We simulate the deanonymization attack with extensive experiments on three real-world datasets, and the experimental results demonstrate that our approach can improve the accuracy and time complexity of deanonymization compared with the state of the art.