Training and fine-tuning large-scale language models largely benefit from collaborative learning, but the approach has been proven vulnerable to gradient inversion attacks (GIAs), which allow adversaries to reconstruct private training data from shared gradients. Existing defenses mainly employ gradient perturbation techniques, e.g., noise injection or gradient pruning, to disrupt GIAs' direct mapping from gradient space to token space. However, these methods often fall short due to the retention of semantics similarity across gradient, embedding, and token spaces. Attackers can map proximate gradients into similar embeddings, and subsequently correspond them to tokens with similar semantics.

Large-scale models are typically adapted to meet the diverse requirements of model owners and users.However, maintaining multiple specialized versions of the model is inefficient.In response, we propose Aim, a novel model modulation paradigm that enables a single model to exhibit diverse behaviors to meet the specific end requirements.Aim enables two key modulation modes: utility and focus modulations.The former provides model owners with dynamic control over output quality to deliver varying utility levels, and the latter offers users precise control to shift model's focused input features.Aim introduces a logits redistribution strategy that operates in a training data-agnostic and retraining-free manner.We establish a formal foundation to ensure Aim's regulation capability, based on the statistical properties of logits ordering via joint probability distributions.Our evaluation confirms Aim's practicality and versatility for AI model modulation, with tasks spanning image classification, semantic segmentation and text generation, and prevalent architectures including ResNet, SegFormer and Llama.

The wide adoption of deep learning in safety-critical domains has driven the need for formally verifying the robustness of neural networks. A critical challenge in this endeavor lies in addressing the inherent non-linearity of activation functions. The convex hull of the activation function has emerged as a promising solution, as it effectively tightens variable ranges and provides multi-neuron constraints, which together enhance verification precision. Given that constructing exact convex hulls is computationally expensive and even infeasible in most cases, existing research has focused on over-approximating them. Several ad-hoc methods have been devised for specific functions such as ReLU and Sigmoid. Nonetheless, there remains a substantial gap in developing broadly applicable approaches for general activation functions. In this work, we propose WraAct , an approach to efficiently constructing tight over-approximations for activation function hulls. Its core idea is to introduce linear constraints to smooth out the fluctuations in the target function, by leveraging double-linear-piece (DLP) functions to simplify the local geometry. In this way, the problem is reduced to over-approximating DLP functions, which can be efficiently handled. We evaluate WraAct against SBLM+PDDM, the state-of-the-art (SOTA) multi-neuron over-approximation method based on decomposing functions into segments. WraAct outperforms it on commonly-used functions like Sigmoid, Tanh, and MaxPool, offering superior efficiency (average 400X faster) and precision (average 150X) while constructing fewer constraints (average 50% reduction). It can complete the computation of up to 8 input dimensions in 10 seconds. We also integrate WraAct into a neural network verifier to evaluate its capability in verification tasks. On 100 benchmark samples, it significantly enhances the single-neuron verification from under 10 to over 40, and outperforms the multi-neuron verifier PRIMA with up to additional 20 verified samples. On large networks like ResNets with 22k neurons, it can complete the verification of one sample within one minute.

The gradient inversion attack has been demonstrated as a significant privacy threat to federated learning (FL), particularly in continuous domains such as vision models. In contrast, it is often considered less effective or highly dependent on impractical training settings when applied to language models, due to the challenges posed by the discrete nature of tokens in text data. As a result, its potential privacy threats remain largely underestimated, despite FL being an emerging training method for language models. In this work, we propose a domain-specific gradient inversion attack named GRAB (gradient inversion with hybrid optimization). GRAB features two alternating optimization processes to address the challenges caused by practical training settings, including a simultaneous optimization on dropout masks between layers for improved token recovery and a discrete optimization for effective token sequencing. GRAB can recover a significant portion (up to 92.9% recovery rate) of the private training data, outperforming the attack strategy of utilizing discrete optimization with an auxiliary model by notable improvements of up to 28.9% recovery rate in benchmark settings and 48.5% recovery rate in practical settings. GRAB provides a valuable step forward in understanding this privacy threat in the emerging FL training mode of language models.

The growing complexity of deep neural network models in modern application domains necessitates a complex training process that involves extensive data, sophisticated design, and substantial computation. The trained model inherently encapsulates the intellectual property owned by the model developer (or the model owner). Consequently, safeguarding the model from unauthorized use by entities who obtain access to the model (or the model controllers), i.e., preserving the fundamental rights and proprietary interests of the model owner, has become a critical necessity.In this work, we propose CORELOCKER, employing the strategic extraction of a small subset of significant weights from the neural network. This subset serves as the access key to unlock the model’s complete capability. The extraction of the key can be customized to varying levels of utility that the model owner intends to release. Authorized users with the access key have full access to the model, while unauthorized users can have access to only part of its capability. We establish a formal foundation to underpin CORELOCKER, which provides crucial lower and upper bounds for the utility disparity between pre- and post-protected networks. We evaluate CORELOCKER using representative datasets such as Fashion-MNIST, CIFAR-10, and CIFAR-100, as well as real-world models including Vg-gNet, ResNet, and DenseNet. Our experimental results confirm its efficacy. We also demonstrate CORELOCKER’s resilience against advanced model restoration attacks based on fine-tuning and pruning.

Convex hulls are commonly used to tackle the non-linearity of activation functions in the verification of neural networks. Computing the exact convex hull is a costly task though. In this work, we propose a fast and precise approach to over-approximating the convex hull of the ReLU function (referred to as the ReLU hull ), one of the most used activation functions. Our key insight is to formulate a convex polytope that “wraps” the ReLU hull, by reusing the linear pieces of the ReLU function as the lower faces and constructing upper faces that are adjacent to the lower faces. The upper faces can be efficiently constructed based on the edges and vertices of the lower faces, given that an n -dimensional (or simply n d hereafter) hyperplane can be determined by an ( n - 1 ) d hyperplane and a point outside of it. We implement our approach as WraLU , and evaluate its performance in terms of precision, efficiency, constraint complexity, and scalability. WraLU outperforms existing advanced methods by generating fewer constraints to achieve tighter approximation in less time. It exhibits versatility by effectively addressing arbitrary input polytopes and higher-dimensional cases, which are beyond the capabilities of existing methods. We integrate WraLU into PRIMA, a state-of-the-art neural network verifier, and apply it to verify large-scale ReLU-based neural networks. Our experimental results demonstrate that WraLU achieves a high efficiency without compromising precision. It reduces the number of constraints that need to be solved by the linear programming solver by up to half, while delivering comparable or even superior results compared to the state-of-the-art verifiers.