Although boosting software development performance, large language model (LLM)-powered code generation introduces intellectual property and data security risks rooted in the fact that a service provider (cloud) observes a client's prompts and generated code, which can be proprietary in commercial systems. To mitigate this problem, we propose NOIR, the first framework to protect the client's prompts and generated code from the cloud. NOIR uses an encoder and a decoder at the client to encode and send the prompts' embeddings to the cloud to get enriched embeddings from the LLM, which are then decoded to generate the code locally at the client. Since the cloud can use the embeddings to infer the prompt and the generated code, NOIR introduces a new mechanism to achieve indistinguishability, a local differential privacy protection at the token embedding level, in the vocabulary used in the prompts and code, and a data-independent and randomized tokenizer on the client side. These components effectively defend against reconstruction and frequency analysis attacks by an honest-but-curious cloud. Extensive analysis and results using open-source LLMs show that NOIR significantly outperforms existing baselines on benchmarks, including the Evalplus (MBPP and HumanEval, Pass@1 of 76.7 and 77.4), and BigCodeBench (Pass@1 of 38.7, only a 1.77% drop from the original LLM) under strong privacy against attacks.

Adversarial weight perturbation has emerged as a concerning threat to LLMs that either use training privileges or system-level access to inject adversarial corruption in model weights. With the emergence of innovative defensive solutions that place system- and algorithm-level checks and corrections in the input and weight spaces, these perturbations are increasingly susceptible to defenses. This work develops a novel perspective on Trojan attacks that generates an attacker-designed model output while leaving no attack traces on the inputs or weights. Such an attack space can be unlocked through corruption of the key-value (KV) cache. In this paper, we introduce CacheTrap, a novel Trojan attack that corrupts the value vectors stored in the KV cache. These vectors capture the dynamic activations for specific token positions and therefore constitute a natural surface for transient, inference-time trigger insertion. The transient nature of these KV values and their dependence on victim input imply additional constraints on our attack, such as a lack of knowledge of the victim's data or domain application, and, consequently, a lack of gradient information. The objective of the proposed CacheTrap is to develop a vulnerable KV bit-searching algorithm so that, once the attack employs the identified bit-flip as a trigger, the model generates targeted behavior, e.g., classifying inputs towards the target class. Moreover, CacheTrap is a data- and gradient-free attack which also has no impact on the model's utility. Our evaluation demonstrates that the proposed attack enables the first successful Trojan attack on LLMs with a single bit flip in the KV cache. In addition, the data-independent nature of the attack ensures that once the attacker identifies the vulnerable bit index, the location remains constant and can be transferred to a wide range of victim tasks/datasets/queries with no overhead.

Fine-tuning Large Language Models (LLMs) for downstream tasks often compromises safety alignment, even when using parameter-efficient methods like LoRA. In this work, we uncover a notable property: fine-tuned models preserve the geometric structure of their loss landscapes concerning harmful content, regardless of the fine-tuning method employed. This suggests that safety behaviors are not erased but shifted to less influential regions of the parameter space. Building on this insight, we propose a curvature-aware alignment restoration method that leverages influence functions and second-order optimization to selectively increase loss on harmful inputs while preserving task performance. By navigating the shared geometry between base and fine-tuned models, our method discourages unsafe outputs while preserving task-relevant performance, avoiding full reversion and enabling precise, low-impact updates. Extensive evaluations across multiple model families and adversarial settings show that our approach efficiently reduces harmful responses while maintaining or even improving utility and few-shot learning performance.

Large language models (LLMs) have achieved remarkable progress across diverse tasks, yet their internal mechanisms remain largely opaque. In this work, we address a fundamental question: to what extent can the original input text be recovered from a single last-token representation within an LLM? We propose Rep2Text, a novel framework for decoding full text from last-token representations. Rep2Text employs a trainable adapter that projects a target model's internal representations into the embedding space of a decoding language model, which then autoregressively reconstructs the input text. Experiments on various model combinations (Llama-3.1-8B, Gemma-7B, Mistral-7B-v0.1, Llama-3.2-3B) demonstrate that, on average, over half of the information in 16-token sequences can be recovered from this compressed representation while maintaining strong semantic integrity and coherence. Furthermore, our analysis reveals an information bottleneck effect: longer sequences exhibit decreased token-level recovery while preserving strong semantic integrity. Besides, our framework also demonstrates robust generalization to out-of-distribution medical data.

Large language models (LLMs) demonstrate remarkable capabilities across various tasks. However, their deployment introduces significant risks related to intellectual property. In this context, we focus on model stealing attacks, where adversaries replicate the behaviors of these models to steal services. These attacks are highly relevant to proprietary LLMs and pose serious threats to revenue and financial stability. To mitigate these risks, the watermarking solution embeds imperceptible patterns in LLM outputs, enabling model traceability and intellectual property verification. In this paper, we study the vulnerability of LLM service providers by introducing $δ$-STEAL, a novel model stealing attack that bypasses the service provider's watermark detectors while preserving the adversary's model utility. $δ$-STEAL injects noise into the token embeddings of the adversary's model during fine-tuning in a way that satisfies local differential privacy (LDP) guarantees. The adversary queries the service provider's model to collect outputs and form input-output training pairs. By applying LDP-preserving noise to these pairs, $δ$-STEAL obfuscates watermark signals, making it difficult for the service provider to determine whether its outputs were used, thereby preventing claims of model theft. Our experiments show that $δ$-STEAL with lightweight modifications achieves attack success rates of up to $96.95\%$ without significantly compromising the adversary's model utility. The noise scale in LDP controls the trade-off between attack effectiveness and model utility. This poses a significant risk, as even robust watermarks can be bypassed, allowing adversaries to deceive watermark detectors and undermine current intellectual property protection methods.

Ensuring safety is a foundational requirement for large language models (LLMs). Achieving an appropriate balance between enhancing the utility of model outputs and mitigating their potential for harm is a complex and persistent challenge. Contemporary approaches frequently formalize this problem within the framework of Constrained Markov Decision Processes (CMDPs) and employ established CMDP optimization techniques. However, these methods exhibit two notable limitations. First, their reliance on reward and cost functions renders performance highly sensitive to the underlying scoring mechanism, which must capture semantic meaning rather than being triggered by superficial keywords. Second, CMDP-based training entails tuning dual-variable, a process that is both computationally expensive and does not provide any provable safety guarantee for a fixed dual variable that can be exploitable through adversarial jailbreaks. To overcome these limitations, we introduce Certifiable Safe-RLHF (CS-RLHF) that introduces a cost model trained on a large-scale corpus to assign semantically grounded safety scores. In contrast to the lagrangian-based approach, CS-RLHF adopts a rectified penalty-based formulation. This design draws on the theory of exact penalty functions in constrained optimization, wherein constraint satisfaction is enforced directly through a suitably chosen penalty term. With an appropriately scaled penalty, feasibility of the safety constraints can be guaranteed at the optimizer, eliminating the need for dual-variable updates. Empirical evaluation demonstrates that CS-RLHF outperforms state-of-the-art LLM model responses rendering at-least 5 times efficient against nominal and jail-breaking prompts

The rapid rise of synthetic media has made deepfake detection a critical challenge for online safety and trust. Progress remains constrained by the scarcity of large, high-quality datasets. Although multimodal large language models (LLMs) exhibit strong reasoning capabilities, their performance on deepfake detection is poor, often producing explanations that are misaligned with visual evidence or hallucinatory. To address this limitation, we introduce a reasoning-annotated dataset for deepfake detection and propose Paragraph-level Relative Policy Optimization (PRPO), a reinforcement learning algorithm that aligns LLM reasoning with image content at the paragraph level. Experiments show that PRPO improves detection accuracy by a wide margin and achieves the highest reasoning score of 4.55/5.0. Ablation studies further demonstrate that PRPO significantly outperforms GRPO under test-time conditions. These results underscore the importance of grounding multimodal reasoning in visual evidence to enable more reliable and interpretable deepfake detection.

Neural speech synthesis techniques have enabled highly realistic speech deepfakes, posing major security risks. Speech deepfake detection is challenging due to distribution shifts across spoofing methods and variability in speakers, channels, and recording conditions. We explore learning shared discriminative features as a path to robust detection and propose Information Bottleneck enhanced Confidence-Aware Adversarial Network (IB-CAAN). Confidence-guided adversarial alignment adaptively suppresses attack-specific artifacts without erasing discriminative cues, while the information bottleneck removes nuisance variability to preserve transferable features. Experiments on ASVspoof 2019/2021, ASVspoof 5, and In-the-Wild demonstrate that IB-CAAN consistently outperforms baseline and achieves state-of-the-art performance on many benchmarks.

We present CryptGNN, a secure and effective inference solution for third-party graph neural network (GNN) models in the cloud, which are accessed by clients as ML as a service (MLaaS). The main novelty of CryptGNN is its secure message passing and feature transformation layers using distributed secure multi-party computation (SMPC) techniques. CryptGNN protects the client's input data and graph structure from the cloud provider and the third-party model owner, and it protects the model parameters from the cloud provider and the clients. CryptGNN works with any number of SMPC parties, does not require a trusted server, and is provably secure even if P-1 out of P parties in the cloud collude. Theoretical analysis and empirical experiments demonstrate the security and efficiency of CryptGNN.

Federated Learning (FL) can be vulnerable to attacks, such as model poisoning, where adversaries send malicious local weights to compromise the global model. Federated Unlearning (FU) is emerging as a solution to address such vulnerabilities by selectively removing the influence of detected malicious contributors on the global model without complete retraining. However, unlike typical FU scenarios where clients are trusted and cooperative, applying FU with malicious and possibly colluding clients is challenging because their collaboration in unlearning their data cannot be assumed. This work presents FedUP, a lightweight FU algorithm designed to efficiently mitigate malicious clients' influence by pruning specific connections within the attacked model. Our approach achieves efficiency by relying only on clients' weights from the last training round before unlearning to identify which connections to inhibit. Isolating malicious influence is non-trivial due to overlapping updates from benign and malicious clients. FedUP addresses this by carefully selecting and zeroing the highest magnitude weights that diverge the most between the latest updates from benign and malicious clients while preserving benign information. FedUP is evaluated under a strong adversarial threat model, where up to 50%-1 of the clients could be malicious and have full knowledge of the aggregation process. We demonstrate the effectiveness, robustness, and efficiency of our solution through experiments across IID and Non-IID data, under label-flipping and backdoor attacks, and by comparing it with state-of-the-art (SOTA) FU solutions. In all scenarios, FedUP reduces malicious influence, lowering accuracy on malicious data to match that of a model retrained from scratch while preserving performance on benign data. FedUP achieves effective unlearning while consistently being faster and saving storage compared to the SOTA.