ICLR 2026 Papers

We characterize the effectiveness of Sharpness-aware minimization (SAM) under machine unlearning scheme, where unlearning forget signals interferes with learning retain signals. While previous work prove that SAM improves generalization with noise memorization prevention, we show that SAM abandons such denoising property when fitting the forget set, leading to altered generalization depending on signal strength. We further characterize the signal surplus of SAM in the order of signal strength, which enables learning from less retain signals to maintain model performance and putting more weight on unlearning the forget set. Empirical studies show that SAM outperforms SGD with relaxed requirement for retain signals and can enhance various unlearning methods either as pretrain or unlearn algorithm. Motivated by our refined characterization of SAM unlearning and observing that overfitting can benefit more stringent sample-specific unlearning, we propose Sharp MinMax, which splits the model into two to learn retain signals with SAM and unlearn forget signals with sharpness maximization, achieving best performance. Extensive experiments show that SAM enhances unlearning across varying difficulties measured by memorization, yielding decreased feature entanglement between retain and forget sets, stronger resistance to membership inference attacks, and a flatter loss landscape. Our observations generalize to more noised data, different optimizers, and different architectures.

We study the classic problem of prediction with expert advice under the constraint of local differential privacy (LDP). In this context, we first show that a classical algorithm naturally satisfies LDP and then design two new algorithms that improve it: RW-AdaBatch and RW-Meta. For RW-AdaBatch, we exploit the limited-switching behavior induced by LDP to provide a novel form of privacy amplification that grows stronger on easier data, analogous to the shuffle model in offline learning. Drawing on the theory of random walks, we prove that this improvement carries essentially no utility cost. For RW-Meta, we develop a general method for privately selecting between experts that are themselves non-trivial learning algorithms, and we show that in the context of LDP this carries no extra privacy cost. In contrast, prior work has only considered data-independent experts. We also derive formal regret bounds that scale inversely with the degree of independence between experts. Our analysis is supplemented by evaluation on real-world data reported by hospitals during the COVID-19 pandemic; RW-Meta outperforms both the classical baseline and a state-of-the-art \textit{central} DP algorithm by 1.5-3$\times$ on the task of predicting which hospital will report the highest density of COVID patients each week.

Deep neural networks remain highly vulnerable to adversarial examples, and most defenses collapse once gradients can be reliably estimated. We identify \emph{gradient consensus}—the tendency of randomized transformations to yield aligned gradients—as a key driver of adversarial transferability. Attackers exploit this consensus to construct perturbations that remain effective across transformations. We introduce \textbf{DRIFT} (Divergent Response in Filtered Transformations), a stochastic ensemble of lightweight, learnable filters trained to actively disrupt gradient consensus. Unlike prior randomized defenses that rely on gradient masking, DRIFT enforces \emph{gradient dissonance} by maximizing divergence in Jacobian- and logit-space responses while preserving natural predictions. Our contributions are threefold: (i) we formalize gradient consensus and provide a theoretical analysis linking consensus to transferability; (ii) we propose a consensus-divergence training strategy combining prediction consistency, Jacobian separation, logit-space separation, and adversarial robustness; and (iii) we show that DRIFT achieves substantial robustness gains on ImageNet across CNNs and Vision Transformers, outperforming state-of-the-art preprocessing, adversarial training, and diffusion-based defenses under adaptive white-box, transfer-based, and gradient-free attacks. DRIFT delivers these improvements with negligible runtime and memory cost, establishing gradient divergence as a practical and generalizable principle for adversarial defense.

Large Language Models (LLMs) have revolutionized conversational AI, yet their robustness in extended multi-turn dialogues remains poorly understood. Existing evaluation frameworks focus on static benchmarks and single-turn assessments, failing to capture the temporal dynamics of conversational degradation that characterize real-world interactions. In this work, we present a large-scale survival analysis of conversational robustness, modeling failure as a time-to-event process over 36,951 turns from 9 state-of-the-art LLMs on the MT-Consistency benchmark. Our framework combines Cox proportional hazards, Accelerated Failure Time (AFT), and Random Survival Forest models with simple semantic drift features. We find that abrupt prompt-to-prompt semantic drift sharply increases the hazard of inconsistency, whereas cumulative drift is counterintuitively \emph{protective}, suggesting adaptation in conversations that survive multiple shifts. AFT models with model–drift interactions achieve the best combination of discrimination and calibration, and proportional hazards checks reveal systematic violations for key drift covariates, explaining the limitations of Cox-style modeling in this setting. Finally, we show that a lightweight AFT model can be turned into a turn-level risk monitor that flags most failing conversations several turns before the first inconsistent answer while keeping false alerts modest. These results establish survival analysis as a powerful paradigm for evaluating multi-turn robustness and for designing practical safeguards for conversational AI systems.

Adversarial training (AT) is an effective defense for large language models (LLMs) against jailbreak attacks, but performing AT on LLMs is costly. To improve the efficiency of AT for LLMs, recent studies propose continuous AT (CAT) that searches for adversarial inputs within the continuous embedding space of LLMs during AT. While CAT has achieved empirical success, its underlying mechanism, i.e., why adversarial perturbations in the embedding space can help LLMs defend against jailbreak prompts synthesized in the input token space, remains unknown. This paper presents the first theoretical analysis of CAT on LLMs based on in-context learning (ICL) theory. For linear transformers trained with adversarial examples from the embedding space on in-context linear regression tasks, we prove a robust generalization bound that has a negative correlation with the perturbation radius in the embedding space. This clearly explains why CAT can defend against jailbreak prompts from the LLM's token space. Further, the robust bound shows that the robustness of an adversarially trained LLM is closely related to the singular values of its embedding matrix. Based on this, we propose to improve LLM CAT by introducing an additional regularization term, which depends on singular values of the LLM's embedding matrix, into the objective function of CAT. Experiments on real-world LLMs demonstrate that our method can help LLMs achieve better jailbreak robustness-utility tradeoff.

The memorization of training data by neural networks raises pressing concerns for privacy and security. Recent work has shown that, under certain conditions, portions of the training set can be reconstructed directly from model parameters. Some of these methods exploit implicit bias toward margin maximization, suggesting that properties often regarded as beneficial for generalization may actually compromise privacy. Yet despite striking empirical demonstrations, the reliability of these attacks remains poorly understood and lacks a solid theoretical foundation. In this work, we take a complementary perspective: rather than designing stronger attacks, we analyze the inherent weaknesses and limitations of existing reconstruction methods and identify conditions under which they fail. We rigorously prove that, without incorporating prior knowledge about the data, there exist infinitely many alternative solutions that may lie arbitrarily far from the true training set, rendering reconstruction fundamentally unreliable. Empirically, we further demonstrate that exact duplication of training examples occurs only by chance. Our results refine the theoretical understanding of when training set leakage is possible and offer new insights into mitigating reconstruction attacks. Remarkably, we demonstrate that networks trained more extensively, and therefore satisfying implicit bias conditions more strongly -- are, in fact, less susceptible to reconstruction attacks, reconciling privacy with the need for strong generalization in this setting.

As diffusion model-based services expand across various domains, safeguarding client data privacy has become increasingly critical. While fully homomorphic encryption and secure multi-party computation enable privacy-preserving inference, their high computational overhead poses challenges for large-scale diffusion applications. Recent work alleviates computational costs by substituting non-linear operations with low-degree polynomial approximations. While such relaxations reduce latency, they incur significant degradation in generative fidelity, and inference remains considerably slower than plaintext execution. To further accelerate secure inference while preserving performance, we explore more relaxed approximations and propose a score-correction framework that rectifies the conditional score shift induced by the relaxed approximation, rather than decreasing the approximation error itself. The key insight is that unconditional generation can be executed without approximation and thus provides a high-fidelity score signal. Leveraging this unconditional score as corrective guidance enables more relaxed approximations while preserving semantic and perceptual quality. In experiments, we demonstrate that our method significantly alleviates the performance degradation caused by relaxed approximations across various benchmarks.

Multi-turn jailbreaks capture the real threat model for safety-aligned chatbots, where single-turn attacks are merely a special case. Yet existing approaches break under exploration complexity and intent drift. We propose SEMA, a simple yet effective framework that trains a multi-turn attacker without relying on any existing strategies or external data. SEMA comprises two stages. Prefilling self-tuning enables usable rollouts by fine-tuning on non-refusal, well-structured, multi-turn adversarial prompts that are self-generated with a minimal prefix, thereby stabilizing subsequent learning. Reinforcement learning with intent-drift-aware reward trains the attacker to elicit valid multi-turn adversarial prompts while maintaining the same harmful objective. We anchor harmful intent in multi-turn jailbreaks via an intent-drift-aware reward that combines intent alignment, compliance risk, and level of detail. Our open-loop attack regime avoids dependence on victim feedback, unifies single- and multi-turn settings, and reduces exploration complexity. Across multiple datasets, victim models, and jailbreak judges, our method achieves state-of-the-art (SOTA) attack success rates (ASR), outperforming all single-turn baselines, manually scripted and template-driven multi-turn baselines, as well as our SFT (Supervised Fine-Tuning) and DPO (Direct Preference Optimization) variants. For instance, SEMA performs an average 80.1% ASR@1 across three closed-source and open-source victim models on AdvBench, 33.9% over SOTA. The approach is compact, reproducible, and transfers across targets, providing a stronger and more realistic stress test for large language model (LLM) safety and enabling automatic redteaming to expose and localize failure modes.

Multi-modal large reasoning models (MLRMs) pose significant privacy risks by inferring precise geographic locations from personal images through hierarchical chain-of-thought reasoning. Existing privacy protection techniques, primarily designed for perception-based models, prove ineffective against MLRMs' sophisticated multi-step reasoning processes that analyze environmental cues. We introduce **ReasonBreak**, a novel adversarial framework specifically designed to disrupt hierarchical reasoning in MLRMs through concept-aware perturbations. Our approach is founded on the key insight that effective disruption of geographic reasoning requires perturbations aligned with conceptual hierarchies rather than uniform noise. ReasonBreak strategically targets critical conceptual dependencies within reasoning chains, generating perturbations that invalidate specific inference steps and cascade through subsequent reasoning stages. To facilitate this approach, we contribute **GeoPrivacy-6K**, a comprehensive dataset comprising 6,341 ultra-high-resolution images ($\geq$2K) with hierarchical concept annotations. Extensive evaluation across seven state-of-the-art MLRMs (including GPT-o3, GPT-5, Gemini 2.5 Pro) demonstrates ReasonBreak's superior effectiveness, achieving a 14.4\% improvement in tract-level protection (33.8\% vs 19.4\%) and nearly doubling block-level protection (33.5\% vs 16.8\%). This work establishes a new paradigm for privacy protection against reasoning-based threats.

Low-Rank Adaptation (LoRA) benefits from its plug-and-play nature, enabling large language models (LLMs) to achieve significant performance gains at low cost, has driven the development of LoRA-sharing platforms. However, the jailbreak and backdoor concerns associated with LoRA-sharing platforms remain underexplored. Existing LoRA-based attacks primarily focus on achieving high attack success rates, while neglecting the core reason why LoRA is adopted by user, i.e. to gain downstream task capabilities. However, achieving effective attacks while preserving strong multi-task performance remains challenging, as the largely unrelated objectives tend to interfere with each other during optimization. In this paper, we propose JailbreakLoRA, a multi-task jailbreak LoRA training method that balances task utility and attack capability, it resolves training interference by uncertainty-weighting losses and mitigating gradient conflicts. Additionally, JailbreakLoRA is designed to generate an affirmative prefix upon trigger activation, exploiting inference-time hallucinations to enhance the effectiveness of jailbreak. Experimental results demonstrate that our method outperforms SOTA LoRA-based attacks, achieving a 16.0\% improvement in attack success rate while also enhancing performance on multi-downstream tasks by 16.5\% in average.

Machine learning models are known to leak sensitive information, as they inevitably memorize (parts of) their training data. More alarmingly, large language models (LLMs) are now trained on nearly all available data, which amplifies the magnitude of information leakage and raises serious privacy risks. Hence, it is more crucial than ever to quantify privacy risk before the release of LLMs. The standard method to quantify privacy is via membership inference attacks, where the state-of-the-art approach is the Robust Membership Inference Attack (RMIA). In this paper, we present InfoRMIA, a principled information-theoretic formulation of membership inference. Our method consistently outperforms RMIA across benchmarks while also offering improved computational efficiency. In the second part of the paper, we identify the limitations of treating sequence-level membership inference as the gold standard for measuring leakage. We propose a new perspective for studying membership and memorization in LLMs: token-level signals and analyses. We show that a simple token-based InfoRMIA can pinpoint which tokens are memorized within generated outputs, thereby localizing leakage from the sequence level down to individual tokens, while achieving stronger sequence-level inference power on LLMs. This new scope rethinks privacy in LLMs and can lead to more targeted mitigation, such as exact unlearning.

Approximate machine unlearning aims to efficiently remove the influence of specific data points from a trained model, offering a practical alternative to full retraining. However, it introduces privacy risks: an adversary with access to both the original and unlearned models can exploit their differences for membership inference or data reconstruction. We show these vulnerabilities arise from two factors: large gradient norms of forgotten samples and the close proximity of the unlearned model to the original. To demonstrate their severity, we design unlearning-specific membership inference and reconstruction attacks, showing that several state-of-the-art methods (such as NGP and SCRUB) remain vulnerable. To mitigate this leakage, we introduce WARP, a plug-and-play teleportation defense that leverages neural network symmetries to reduce gradient energy of forgotten samples and increase parameter dispersion while preserving accuracy. This reparameterization hides the signal of forgotten data, making it harder for attackers to distinguish forgotten samples from non-members or to recover them through reconstruction. Across six unlearning algorithms, our approach achieves consistent privacy gains, reducing adversarial advantage by up to 64% in black-box settings and 92% in white-box settings, while maintaining accuracy on retained data. These results highlight teleportation as a general tool for improving privacy in approximate unlearning.

Concept erasure helps stop diffusion models (DMs) from generating harmful content; but current methods face robustness-retention trade-off. **Robustness** means the model fine-tuned by concept erasure methods resists reactivation of erased concepts, even under semantically related prompts. **Retention** means unrelated concepts are preserved so the model’s overall utility stays intact. Both are critical for concept erasure in practice, yet addressing them simultaneously is challenging, as existing works typically improve one factor while sacrificing the other. Prior work typically strengthens one while degrading the other—e.g., mapping a single erased prompt to a fixed safe target leaves class-level remnants exploitable by prompt attacks, whereas retention-oriented schemes underperform against adaptive adversaries. This paper introduces AEGIS (Adversarial Erasure with Gradient-Informed Synergy), a retention-data-free framework that advances both robustness and retention. First, AEGIS replaces handpicked targets with an Adversarial Erasure Target (AET) optimized to approximate the semantic center of the erased concept class. By aligning the model’s prediction on the erased prompt to an AET-derived target in the shared text–image space, AEGIS increases predicted-noise distances not just for the instance but for semantically related variants, substantially hardening the DMs against state-of-the-art adversarial prompt attacks. Second, AEGIS preserves utility without auxiliary data via Gradient Regularization Projection (GRP), a conflict-aware gradient rectification that selectively projects away the destructive component of the retention update only when it opposes the erasure direction. This directional, data-free projection mitigates interference between erasure and retention, avoiding dataset bias and accidental relearning. Extensive experiments show that AEGIS markedly reduces attack success rates across various concepts while maintaining or improving FID/CLIP versus advanced baselines, effectively pushing beyond the prevailing robustness–retention trade-off. The source code is in the supplementary.

In this work we study black-box privacy auditing, where the goal is to lower bound the privacy parameter of a differentially private learning algorithm using only the algorithm’s outputs (i.e., final trained model). For DP-SGD (the most successful method for training differentially private deep learning models), the canonical approach auditing uses membership inference—an auditor comes with a small set of special “ca- nary” examples, inserts a random subset of them into the training set, and then tries to discern which of their canaries were included in the training set (typically via a membership inference attack). The audi- tor’s success rate then provides a lower bound on the privacy parameters of the learning algorithm. Our main contribution is a method for optimizing the auditor’s canary set to improve privacy auditing, leverag- ing recent work on metagradient optimization. Our empirical evaluation demonstrates that by using such optimized canaries, we can improve empirical lower bounds for differentially private image classification models by over 2x in certain instances. Furthermore, we demonstrate that our method is transferable and efficient: canaries optimized for non-private SGD with a small model architecture remain effective when auditing larger models trained with DP-SGD.

Diffusion models have revolutionized image synthesis but raise concerns around content provenance and authenticity. Digital watermarking offers a means of tracing generated media, yet traditional schemes often introduce distributional shifts and degrade visual quality. Recent lossless methods embed watermark bits directly into the latent Gaussian prior without modifying model weights, but still require per-image key storage or heavy cryptographic overhead. In this paper, we introduce Spherical Watermark, an encryption‐free and lossless watermarking framework that integrates seamlessly with diffusion architectures. First, our binary embedding module mixes repeated watermark bits with random padding to form a high-entropy code. Second, the spherical mapping module projects this code onto the unit sphere, applies an orthogonal rotation, and scales by a chi-square-distributed radius to recover exact multivariate Gaussian noise. We theoretically prove that the watermarked noise distribution preserves the target prior up to third-order moments, and empirically demonstrate that it is statistically indistinguishable from a standard multivariate normal distribution. Adopting Stable Diffusion, extensive experiments confirm that Spherical Watermark consistently preserves high visual fidelity while simultaneously improving traceability, computational efficiency, and robustness under attacks, thereby outperforming both lossy and lossless approaches.

As LLMs are increasingly deployed in real-world applications, ensuring their ability to refuse malicious prompts, especially jailbreak attacks, is essential for safe and reliable use. Recently, activation steering has emerged as an effective approach for enhancing LLM safety by adding a refusal direction vector to internal activations of LLMs during inference, which will further induce the refusal behaviors of LLMs. However, indiscriminately applying activation steering fundamentally suffers from the trade-off between safety and utility, since the same steering vector can also lead to over-refusal and degraded performance on benign prompts. Although prior efforts, such as vector calibration and conditional steering, have attempted to mitigate this trade-off, their lack of theoretical grounding limits their robustness and effectiveness. To better address the trade-off between safety and utility, we present a theoretically grounded and empirically effective activation steering method called AlphaSteer. Specifically, it considers activation steering as a learnable process with two principled learning objectives: utility preservation and safety enhancement. For utility preservation, it learns to construct a nearly zero vector for steering benign data, with the null-space constraints. For safety enhancement, it learns to construct a refusal direction vector for steering malicious data, with the help of linear regression. Experiments across multiple jailbreak attacks and utility benchmarks demonstrate the effectiveness of AlphaSteer, which significantly improves the safety of LLMs without compromising their general capabilities. Our codes are available at \url{https://anonymous.4open.science/r/AlphaSteer-929C/}.

Generative models can generate photorealistic images at scale. This raises urgent concerns about the ability to detect synthetically generated images and attribute these images to specific sources. While watermarking has emerged as a possible solution, existing methods remain fragile to realistic distortions, susceptible to adaptive removal, and expensive to update when the underlying watermarking key changes. We propose a general watermarking framework that formulates the encoding problem as key-dependent perturbation of the parameters of a generative model. Within this framework, we introduce Mixture of LoRA Markers (MOLM), a routing-based instantiation in which binary keys activate lightweight LoRA adapters inside residual and attention blocks. This design avoids key-specific re-training and achieves the desired properties such as imperceptibility, fidelity, verifiability, and robustness. Experiments on Stable Diffusion and FLUX show that MOLM preserves image quality while achieving robust key recovery against distortions, compression and regeneration, averaging attacks, and black-box adversarial attacks on the extractor.

Federated Learning (FL) was initially proposed as a privacy-preserving machine learning paradigm. However, FL has been shown to be susceptible to a series of privacy attacks. Recently, there has been concern about the Source Inference Attack (SIA), where an honest-but-curious central server attempts to identify exactly which client owns a given data point which was used in the training phase. Alarmingly, standard gradient obfuscation techniques with Differential Privacy have been shown to be ineffective against SIAs, at least without severely diminishing the accuracy. In this work, we propose a defense against SIAs within the widely studied shuffle model of FL, where an honest shuffler acts as an intermediary between the clients and the server. First, we demonstrate that standard naive shuffling alone is insufficient to prevent SIAs. To effectively defend against SIAs, shuffling needs to be applied at a more granular level; we propose a novel combination of parameter-level shuffling with the residue number system (RNS). Our approach provides robust protection against SIAs without affecting the accuracy of the joint model and can be seamlessly integrated into other privacy protection mechanisms. We conduct experiments on a series of models and datasets, confirming that standard shuffling approaches fail to prevent SIAs and that, in contrast, our proposed method reduce the attack’s accuracy to the level of random guessing.

Transformer components such as non-linear activations and normalization are inherently non-injective, suggesting that different inputs could map to the same output and prevent exact recovery of the input from a model’s representations. In this paper, we challenge this view. First, we prove mathematically that transformer language models mapping discrete input sequences to their corresponding sequence of continuous representations are injective and therefore lossless, a property established at initialization and preserved during training. Second, we confirm this result empirically through billions of collision tests on six state-of-the-art language models, and observe no collisions. Third, we operationalize injectivity: we introduce SipIt, the first algorithm that provably and efficiently reconstructs the exact input text from hidden activations, establishing linear-time guarantees and demonstrating exact invertibility in practice. Overall, our work establishes injectivity as a fundamental and exploitable property of language models, with direct implications for transparency, interpretability, and safe deployment.