Recent years have seen a surge in research that develops and applies machine learning algorithms to create intelligent learning systems. However, traditional machine learning algorithms have primarily focused on optimizing accuracy and efficiency, and they often fail to consider how to foster trustworthiness in their design. As a result, machine learning models usually face a trust crisis in real-world applications. Driven by these urgent concerns about trustworthiness, in this talk, I will introduce my research efforts towards the goal of making machine learning trustworthy. Specifically, I will delve into the following key research topics: security vulnerabilities and robustness, model explanations, and privacy-preserving mechanisms.
论文检索
输入标题、作者或关键词，从 2,111 篇学术成果中精准定位
已筛选
关键词：Security
2,111篇论文匹配“Security”
第 83 / 106 页
Blind-Touch: Homomorphic Encryption-Based Distributed Neural Network Inference for Privacy-Preserving Fingerprint Authentication
PDF ↗Fingerprint authentication is a popular security mechanism for smartphones and laptops. However, its adoption in web and cloud environments has been limited due to privacy concerns over storing and processing biometric data on servers. This paper introduces Blind-Touch, a novel machine learning-based fingerprint authentication system leveraging homomorphic encryption to address these privacy concerns. Homomorphic encryption allows computations on encrypted data without decrypting. Thus, Blind-Touch can keep fingerprint data encrypted on the server while performing machine learning operations. Blind-Touch combines three strategies to efficiently utilize homomorphic encryption in machine learning: (1) It optimizes the feature vector for a distributed architecture, processing the first fully connected layer (FC-16) in plaintext on the client side and the subsequent layer (FC-1) post-encryption on the server, thereby minimizing encrypted computations; (2) It employs a homomorphic encryption-compatible data compression technique capable of handling 8,192 authentication results concurrently; and (3) It utilizes a clustered server architecture to simultaneously process authentication results, thereby enhancing scalability with increasing user numbers. Blind-Touch achieves high accuracy on two benchmark fingerprint datasets, with a 93.6% F1- score for the PolyU dataset and a 98.2% F1-score for the SOKOTO dataset. Moreover, Blind-Touch can match a fingerprint among 5,000 in about 0.65 seconds. With its privacy-focused design, high accuracy, and efficiency, Blind-Touch is a promising alternative to conventional fingerprint authentication for web and cloud applications.
GaLileo: General Linear Relaxation Framework for Tightening Robustness Certification of Transformers
PDF ↗Transformers based on attention mechanisms exhibit vulnerability to adversarial examples, posing a substantial threat to the security of their applications. Aiming to solve this problem, the concept of robustness certification is introduced to formally ascertain the presence of any adversarial example within a specified region surrounding a given sample. However, prior works have neglected the dependencies among inputs of softmax (the most complex function in attention mechanisms) during linear relaxations. This oversight has consequently led to imprecise certification results. In this work, we introduce GaLileo, a general linear relaxation framework designed to certify the robustness of Transformers. GaLileo effectively surmounts the trade-off between precision and efficiency in robustness certification through our innovative n-dimensional relaxation approach. Notably, our relaxation technique represents a pioneering effort as the first linear relaxation for n-dimensional functions such as softmax. Our novel approach successfully transcends the challenges posed by the curse of dimensionality inherent in linear relaxations, thereby enhancing linear bounds by incorporating input dependencies. Our evaluations encompassed a thorough analysis utilizing the SST and Yelp datasets along with diverse Transformers of different depths and widths. The experimental results demonstrate that, as compared to the baseline method CROWN-BaF, GaLileo achieves up to 3.24 times larger certified radii while requiring similar running times. Additionally, GaLileo successfully attains certification for Transformers' robustness against multi-word lp perturbations, marking a notable accomplishment in this field.
This paper reveals a data bias issue that can profoundly hinder the performance of machine learning models in malicious URL detection. We describe how such bias can be diagnosed using interpretable machine learning techniques and further argue that such biases naturally exist in the real world security data for training a classification model. To counteract these challenges, we propose a debiased training strategy that can be applied to most deep-learning based models to alleviate the negative effects of the biased features. The solution is based on the technique of adversarial training to train deep neural networks learning invariant embedding from biased data. Through extensive experimentation, we substantiate that our innovative strategy fosters superior generalization capabilities across both CNN-based and RNN-based detection models. The findings presented in this work not only expose a latent issue in the field but also provide an actionable remedy, marking a significant step forward in the pursuit of more reliable and robust malicious URL detection.
Warning: this paper contains data, prompts, and model outputs that are offensive in nature. Recently, there has been a surge of interest in integrating vision into Large Language Models (LLMs), exemplified by Visual Language Models (VLMs) such as Flamingo and GPT-4. This paper sheds light on the security and safety implications of this trend. First, we underscore that the continuous and high-dimensional nature of the visual input makes it a weak link against adversarial attacks, representing an expanded attack surface of vision-integrated LLMs. Second, we highlight that the versatility of LLMs also presents visual attackers with a wider array of achievable adversarial objectives, extending the implications of security failures beyond mere misclassification. As an illustration, we present a case study in which we exploit visual adversarial examples to circumvent the safety guardrail of aligned LLMs with integrated vision. Intriguingly, we discover that a single visual adversarial example can universally jailbreak an aligned LLM, compelling it to heed a wide range of harmful instructions (that it otherwise would not) and generate harmful content that transcends the narrow scope of a `few-shot' derogatory corpus initially employed to optimize the adversarial example. Our study underscores the escalating adversarial risks associated with the pursuit of multimodality. Our findings also connect the long-studied adversarial vulnerabilities of neural networks to the nascent field of AI alignment. The presented attack suggests a fundamental adversarial challenge for AI alignment, especially in light of the emerging trend toward multimodality in frontier foundation models.
Notions of transition invariants and closure certificates have seen recent use in the formal verification of controlled dynamical systems against \omega-regular properties. Unfortunately, existing approaches face limitations in two directions. First, they require a closed-form mathematical expression representing the model of the system. Such an expression may be difficult to find, too complex to be of any use, or unavailable due to security or privacy constraints. Second, finding such invariants typically rely on optimization techniques such as sum-of-squares (SOS) or satisfiability modulo theory (SMT) solvers. This restricts the classes of systems that need to be formally verified. To address these drawbacks, we introduce a notion of neural closure certificates. We present a data-driven algorithm that trains a neural network to represent a closure certificate. Our approach is formally correct under some mild assumptions, i.e., one is able to formally show that the unknown system satisfies the \omega-regular property of interest if a neural closure certificate can be computed. Finally, we demonstrate the efficacy of our approach with relevant case studies.
SentinelLMs: Encrypted Input Adaptation and Fine-Tuning of Language Models for Private and Secure Inference
PDF ↗This paper addresses the privacy and security concerns associated with deep neural language models, which serve as crucial components in various modern AI-based applications. These models are often used after being pre-trained and fine-tuned for specific tasks, with deployment on servers accessed through the internet. However, this introduces two fundamental risks: (a) the transmission of user inputs to the server via the network gives rise to interception vulnerabilities, and (b) privacy concerns emerge as organizations that deploy such models store user data with restricted context. To address this, we propose a novel method to adapt and fine-tune transformer-based language models on passkey-encrypted user-specific text. The original pre-trained language model first undergoes a quick adaptation (without any further pre-training) with a series of irreversible transformations applied to the tokenizer and token embeddings. This enables the model to perform inference on encrypted inputs while preventing reverse engineering of text from model parameters and intermediate outputs. After adaptation, models are fine-tuned on encrypted versions of existing training datasets. Experimental evaluation employing adapted versions of renowned models (e.g., BERT, RoBERTa) across established benchmark English and multilingual datasets for text classification and sequence labeling shows that encrypted models achieve performance parity with their original counterparts. This serves to safeguard performance, privacy, and security cohesively.
Revisiting the Information Capacity of Neural Network Watermarks: Upper Bound Estimation and Beyond
PDF ↗To trace the copyright of deep neural networks, an owner can embed its identity information into its model as a watermark. The capacity of the watermark quantify the maximal volume of information that can be verified from the watermarked model. Current studies on capacity focus on the ownership verification accuracy under ordinary removal attacks and fail to capture the relationship between robustness and fidelity. This paper studies the capacity of deep neural network watermarks from an information theoretical perspective. We propose a new definition of deep neural network watermark capacity analogous to channel capacity, analyze its properties, and design an algorithm that yields a tight estimation of its upper bound under adversarial overwriting. We also propose a universal non-invasive method to secure the transmission of the identity message beyond capacity by multiple rounds of ownership verification. Our observations provide evidence for neural network owners and defenders that are curious about the tradeoff between the integrity of their ownership and the performance degradation of their products.
Despite the increasing popularity of graph neural networks (GNNs), the security risks associated with their deployment have not been well explored. Existing works follow the standard adversarial attacks to maximize cross-entropy loss within an L-infinity norm bound. We analyze the robustness of GNNs against node injection attacks (NIAs) in black-box settings by allowing new nodes to be injected and attacked. In this work, we propose to design stronger and transferable NIAs. First, we propose margin aware attack (MAA) that uses a maximum margin loss to generate NIAs. We then propose a novel margin and direction aware attack (MDA) that diversifies the initial directions of MAA attack by minimizing the cosine similarity of the injected nodes with respect to their respective random initialization in addition to the maximization of max-margin loss. This makes the NIAs stronger. We further observe that using L2 norm of gradients in the attack step leads to an enhanced diversity amongst the node features, thereby further enhancing the strength of the attack. We incorporate transferability in NIAs by perturbing the surrogate model before generating the attack. An analysis of eigen spectrum density of the hessian of the loss emphasizes that perturbing the weights of the surrogate model improves the transferability. Our experimental results demonstrate that the proposed resilient node injection attack (R-NIA) consistently outperform PGD by margins about 7-15% on both large and small graph datasets. R-NIA is significantly stronger and transferable than existing NIAs on graph robustness benchmarks.
3D point cloud has been wildly used in security crucial domains, such as self-driving and 3D face recognition. Backdoor attack is a serious threat that usually destroy Deep Neural Networks (DNN) in the training stage. Though a few 3D backdoor attacks are designed to achieve guaranteed attack efficiency, their deformation will alarm human inspection. To obtain invisible backdoored point cloud, this paper proposes a novel 3D backdoor attack, named IBAPC, which generates backdoor trigger in the graph spectral domain. The effectiveness is grounded by the advantage of graph spectral signal that it can induce both global structure and local points to be responsible for the caused deformation in spatial domain. In detail, a new backdoor implanting function is proposed whose aim is to transform point cloud to graph spectral signal for conducting backdoor trigger. Then, we design a backdoor training procedure which updates the parameter of backdoor implanting function and victim 3D DNN alternately. Finally, the backdoored 3D DNN and its associated backdoor implanting function is obtained by finishing the backdoor training procedure. Experiment results suggest that IBAPC achieves SOTA attack stealthiness from three aspects including objective distance measurement, subjective human evaluation, graph spectral signal residual. At the same time, it obtains competitive attack efficiency. The code is available at https://github.com/f-lk/IBAPC.
We propose a combinatorial optimisation model called Limited Query Graph Connectivity Test. We consider a graph whose edges have two possible states (On/Off). The edges' states are hidden initially. We could query an edge to reveal its state. Given a source s and a destination t, we aim to test s−t connectivity by identifying either a path (consisting of only On edges) or a cut (consisting of only Off edges). We are limited to B queries, after which we stop regardless of whether graph connectivity is established. We aim to design a query policy that minimizes the expected number of queries. Our model is mainly motivated by a cyber security use case where we need to establish whether attack paths exist in a given network, between a source (i.e., a compromised user node) and a destination (i.e., a high-privilege admin node). Edge query is resolved by manual effort from the IT admin, which is the motivation behind query minimization. Our model is highly related to Stochastic Boolean Function Evaluation (SBFE). There are two existing exact algorithms for SBFE that are prohibitively expensive. We propose a signifcantly more scalable exact algorithm. While previous exact algorithms only scale for trivial graphs (i.e., past works experimented on at most 20 edges), we empirically demonstrate that our algorithm is scalable for a wide range of much larger practical graphs (i.e., graphs representing Windows domain networks with tens of thousands of edges). We also propose three heuristics. Our best-performing heuristic is via limiting the planning horizon of the exact algorithm. The other two are via reinforcement learning (RL) and Monte Carlo tree search (MCTS). We also derive an algorithm for computing the performance lower bound. Experimentally, we show that all our heuristics are near optimal. The heuristic building on the exact algorithm outperforms all other heuristics, surpassing RL, MCTS and eight existing heuristics ported from SBFE and related literature.
Distributed learning frameworks aim to train global models by sharing gradients among clients while preserving the data privacy of each individual client. However, extensive research has demonstrated that these learning frameworks do not absolutely ensure the privacy, as training data can be reconstructed from shared gradients. Nevertheless, the existing privacy-breaking attack methods have certain limitations. Some are applicable only to small models, while others can only recover images in small batch size and low resolutions, or with low fidelity. Furthermore, when there are some data with the same label in a training batch, existing attack methods usually perform poorly. In this work, we successfully address the limitations of existing attacks by two steps. Firstly, we model the coefficient of variation (CV) of features and design an evolutionary algorithm based on the minimum CV to accurately reconstruct the labels of all training data. After that, we propose a stepwise gradient inversion attack, which dynamically adapts the objective function, thereby effectively and rationally promoting the convergence of attack results towards an optimal solution. With these two steps, our method is able to recover high resolution images (224*224 pixel, from ImageNet and Web) with high fidelity in distributed learning scenarios involving complex models and larger batch size. Experiment results demonstrate the superiority of our approach, reveal the potential vulnerabilities of the distributed learning paradigm, and emphasize the necessity of developing more secure mechanisms. Source code is available at https://github.com/MiLab-HITSZ/2023YeHFGradInv.
While deep learning models have shown significant performance across various domains, their deployment needs extensive resources and advanced computing infrastructure. As a solution, Machine Learning as a Service (MLaaS) has emerged, lowering the barriers for users to release or productize their deep learning models. However, previous studies have highlighted potential privacy and security concerns associated with MLaaS, and one primary threat is model extraction attacks. To address this, there are many defense solutions but they suffer from unrealistic assumptions and generalization issues, making them less practical for reliable protection. Driven by these limitations, we introduce a novel defense mechanism, SAME, based on the concept of sample reconstruction. This strategy imposes minimal prerequisites on the defender's capabilities, eliminating the need for auxiliary Out-of-Distribution (OOD) datasets, user query history, white-box model access, and additional intervention during model training. It is compatible with existing active defense methods. Our extensive experiments corroborate the superior efficacy of SAME over state-of-the-art solutions. Our code is available at https://github.com/xythink/SAME.
Robustness and privacy protection are two important factors of trustworthy federated learning (FL). Existing FL works usually secure data privacy by perturbing local model gradients via the differential privacy (DP) technique, or defend against poisoning attacks by filtering the local gradients in the outlier of the gradient distribution before aggregation. However, these two issues are often addressed independently in existing works, and how to secure federated learning in both privacy and robustness still needs further exploration. In this paper, we unveil that although DP noisy perturbation can improve the learning robustness, DP-FL frameworks are not inherently robust and are vulnerable to a carefully-designed attack method. Furthermore, we reveal that it is challenging for existing robust FL methods to defend against attacks on DP-FL. This can be attributed to the fact that the local gradients of DP-FL are perturbed by random noise, and the selected central gradients inevitably incorporate a higher proportion of poisoned gradients compared to conventional FL. To address this problem, we further propose a new defense method for DP-FL (named Robust-DPFL), which can effectively distinguish poisoned and clean local gradients in DP-FL and robustly update the global model. Experiments on three benchmark datasets demonstrate that baseline methods cannot ensure task accuracy, data privacy, and robustness simultaneously, while Robust-DPFL can effectively enhance the privacy protection and robustness of federated learning meanwhile maintain the task performance.
Model extraction attacks (MEAs) enable an attacker to replicate the functionality of a victim deep neural network (DNN) model by only querying its API service remotely, posing a severe threat to the security and integrity of pay-per-query DNN-based services. Although the majority of current research on MEAs has primarily concentrated on neural classifiers, there is a growing prevalence of image-to-image translation (I2IT) tasks in our everyday activities. However, techniques developed for MEA of DNN classifiers cannot be directly transferred to the case of I2IT, rendering the vulnerability of I2IT models to MEA attacks often underestimated. This paper unveils the threat of MEA in I2IT tasks from a new perspective. Diverging from the traditional approach of bridging the distribution gap between attacker queries and victim training samples, we opt to mitigate the effect caused by the different distributions, known as the domain shift. This is achieved by introducing a new regularization term that penalizes high-frequency noise, and seeking a flatter minimum to avoid overfitting to the shifted distribution. Extensive experiments on different image translation tasks, including image super-resolution and style transfer, are performed on different backbone victim models, and the new design consistently outperforms the baseline by a large margin across all metrics. A few real-life I2IT APIs are also verified to be extremely vulnerable to our attack, emphasizing the need for enhanced defenses and potentially revised API publishing policies.
The field of few-shot learning (FSL) has shown promising results in scenarios where training data is limited, but its vulnerability to backdoor attacks remains largely unexplored. We first explore this topic by first evaluating the performance of the existing backdoor attack methods on few-shot learning scenarios. Unlike in standard supervised learning, existing backdoor attack methods failed to perform an effective attack in FSL due to two main issues. Firstly, the model tends to overfit to either benign features or trigger features, causing a tough trade-off between attack success rate and benign accuracy. Secondly, due to the small number of training samples, the dirty label or visible trigger in the support set can be easily detected by victims, which reduces the stealthiness of attacks. It seemed that FSL could survive from backdoor attacks. However, in this paper, we propose the Few-shot Learning Backdoor Attack (FLBA) to show that FSL can still be vulnerable to backdoor attacks. Specifically, we first generate a trigger to maximize the gap between poisoned and benign features. It enables the model to learn both benign and trigger features, which solves the problem of overfitting. To make it more stealthy, we hide the trigger by optimizing two types of imperceptible perturbation, namely attractive and repulsive perturbation, instead of attaching the trigger directly. Once we obtain the perturbations, we can poison all samples in the benign support set into a hidden poisoned support set and fine-tune the model on it. Our method demonstrates a high Attack Success Rate (ASR) in FSL tasks with different few-shot learning paradigms while preserving clean accuracy and maintaining stealthiness. This study reveals that few-shot learning still suffers from backdoor attacks, and its security should be given attention.
Accurately typing entity mentions from text segments is a fundamental task for various natural language processing applications. Many previous approaches rely on massive human-annotated data to perform entity typing. Nevertheless, collecting such data in highly specialized science and engineering domains (e.g., software engineering and security) can be time-consuming and costly, without mentioning the domain gaps between training and inference data if the model needs to be applied to confidential datasets. In this paper, we study the task of seed-guided fine-grained entity typing in science and engineering domains, which takes the name and a few seed entities for each entity type as the only supervision and aims to classify new entity mentions into both seen and unseen types (i.e., those without seed entities). To solve this problem, we propose SEType which first enriches the weak supervision by finding more entities for each seen type from an unlabeled corpus using the contextualized representations of pre-trained language models. It then matches the enriched entities to unlabeled text to get pseudo-labeled samples and trains a textual entailment model that can make inferences for both seen and unseen types. Extensive experiments on two datasets covering four domains demonstrate the effectiveness of SEType in comparison with various baselines. Code and data are available at: https://github.com/yuzhimanhua/SEType.
Electronic health records (EHRs) have become the foundation of machine learning applications in healthcare, while the utility of real patient records is often limited by privacy and security concerns. Synthetic EHR generation provides an additional perspective to compensate for this limitation. Most existing methods synthesize new records based on real EHR data, without consideration of different types of events in EHR data, which cannot control the event combinations in line with medical common sense. In this paper, we propose MSIC, a Multi-visit health Status Inference model for Collaborative EHR synthesis to address these limitations. First, we formulate the synthetic EHR generation process as a probabilistic graphical model and tightly connect different types of events by modeling the latent health states. Then, we derive a health state inference method tailored for the multi-visit scenario to effectively utilize previous records to synthesize current and future records. Furthermore, we propose to generate medical reports to add textual descriptions for each medical event, providing broader applications for synthesized EHR data. For generating different paragraphs in each visit, we incorporate a multi-generator deliberation framework to collaborate the message passing of multiple generators and employ a two-phase decoding strategy to generate high-quality reports. Our extensive experiments on the widely used benchmarks, MIMIC-III and MIMIC-IV, demonstrate that MSIC advances state-of-the-art results on the quality of synthetic data while maintaining low privacy risks.
Large language models (LLMs) often contain misleading content, emphasizing the need to align them with human values to ensure secure AI systems. Reinforcement learning from human feedback (RLHF) has been employed to achieve this alignment. However, it encompasses two main drawbacks: (1) RLHF exhibits complexity, instability, and sensitivity to hyperparameters in contrast to SFT. (2) Despite massive trial-and-error, multiple sampling is reduced to pair-wise contrast, thus lacking contrasts from a macro perspective. In this paper, we propose Preference Ranking Optimization (PRO) as an efficient SFT algorithm to directly fine-tune LLMs for human alignment. PRO extends the pair-wise contrast to accommodate preference rankings of any length. By iteratively contrasting candidates, PRO instructs the LLM to prioritize the best response while progressively ranking the rest responses. In this manner, PRO effectively transforms human alignment into aligning the probability ranking of n responses generated by LLM with the preference ranking of humans towards these responses. Experiments have shown that PRO outperforms baseline algorithms, achieving comparable results to ChatGPT and human responses through automatic-based, reward-based, GPT-4, and human evaluations.
Out-of-distribution (OOD) detection is a rapidly growing field due to new robustness and security requirements driven by an increased number of AI-based systems. Existing OOD textual detectors often rely on anomaly scores (\textit{e.g.}, Mahalanobis distance) computed on the embedding output of the last layer of the encoder. In this work, we observe that OOD detection performance varies greatly depending on the task and layer output. More importantly, we show that the usual choice (the last layer) is rarely the best one for OOD detection and that far better results can be achieved, provided that an oracle selects the best layer. We propose a data-driven, unsupervised method to leverage this observation to combine layer-wise anomaly scores. In addition, we extend classical textual OOD benchmarks by including classification tasks with a more significant number of classes (up to 150), which reflects more realistic settings. On this augmented benchmark, we show that the proposed post-aggregation methods achieve robust and consistent results comparable to using the best layer according to an oracle while removing manual feature selection altogether.