Search

The standard Reinforcement Learning from Human Feedback (RLHF) framework primarily focuses on optimizing the performance of large language models using pre-collected prompts. However, collecting prompts that provide comprehensive coverage is both tedious and challenging, and often fails to include scenarios that LLMs need to improve on the most. In this paper, we investigate alignment through the lens of two-agent games, involving iterative interactions between an adversarial and a defensive agent. The adversarial agent's task at each step is to generate prompts that expose the weakness of the defensive agent. In return, the defensive agent seeks to improve its responses to these newly identified prompts it struggled with, based on feedback from the reward model. We theoretically demonstrate that this iterative reinforcement learning optimization converges to a Nash Equilibrium for the game induced by the agents. Experimental results in safety scenarios demonstrate that learning in such a competitive environment not only fully trains agents but also leads to policies with enhanced generalization capabilities for both adversarial and defensive agents., Comment: Our code is released at https://github.com/ruizheng20/gpo

We study how to watermark LLM outputs, i.e. embedding algorithmically detectable signals into LLM-generated text to track misuse. Unlike the current mainstream methods that work with a fixed LLM, we expand the watermark design space by including the LLM tuning stage in the watermark pipeline. While prior works focus on token-level watermark that embeds signals into the output, we design a model-level watermark that embeds signals into the LLM weights, and such signals can be detected by a paired detector. We propose a co-training framework based on reinforcement learning that iteratively (1) trains a detector to detect the generated watermarked text and (2) tunes the LLM to generate text easily detectable by the detector while keeping its normal utility. We empirically show that our watermarks are more accurate, robust, and adaptable (to new attacks). It also allows watermarked model open-sourcing. In addition, if used together with alignment, the extra overhead introduced is low - only training an extra reward model (i.e. our detector). We hope our work can bring more effort into studying a broader watermark design that is not limited to working with a fixed LLM. We open-source the code: https://github.com/xiaojunxu/learning-to-watermark-llm .

We explore machine unlearning (MU) in the domain of large language models (LLMs), referred to as LLM unlearning. This initiative aims to eliminate undesirable data influence (e.g., sensitive or illegal information) and the associated model capabilities, while maintaining the integrity of essential knowledge generation and not affecting causally unrelated information. We envision LLM unlearning becoming a pivotal element in the life-cycle management of LLMs, potentially standing as an essential foundation for developing generative AI that is not only safe, secure, and trustworthy, but also resource-efficient without the need of full retraining. We navigate the unlearning landscape in LLMs from conceptual formulation, methodologies, metrics, and applications. In particular, we highlight the often-overlooked aspects of existing LLM unlearning research, e.g., unlearning scope, data-model interaction, and multifaceted efficacy assessment. We also draw connections between LLM unlearning and related areas such as model editing, influence functions, model explanation, adversarial training, and reinforcement learning. Furthermore, we outline an effective assessment framework for LLM unlearning and explore its applications in copyright and privacy safeguards and sociotechnical harm reduction.

We propose and analyze an adaptive adversary that can retrain a Trojaned DNN and is also aware of SOTA output-based Trojaned model detectors. We show that such an adversary can ensure (1) high accuracy on both trigger-embedded and clean samples and (2) bypass detection. Our approach is based on an observation that the high dimensionality of the DNN parameters provides sufficient degrees of freedom to simultaneously achieve these objectives. We also enable SOTA detectors to be adaptive by allowing retraining to recalibrate their parameters, thus modeling a co-evolution of parameters of a Trojaned model and detectors. We then show that this co-evolution can be modeled as an iterative game, and prove that the resulting (optimal) solution of this interactive game leads to the adversary successfully achieving the above objectives. In addition, we provide a greedy algorithm for the adversary to select a minimum number of input samples for embedding triggers. We show that for cross-entropy or log-likelihood loss functions used by the DNNs, the greedy algorithm provides provable guarantees on the needed number of trigger-embedded input samples. Extensive experiments on four diverse datasets -- MNIST, CIFAR-10, CIFAR-100, and SpeechCommand -- reveal that the adversary effectively evades four SOTA output-based Trojaned model detectors: MNTD, NeuralCleanse, STRIP, and TABOR.

Accurately predicting molecular properties is a challenging but essential task in drug discovery. Recently, many mono-modal deep learning methods have been successfully applied to molecular property prediction. However, the inherent limitation of mono-modal learning arises from relying solely on one modality of molecular representation, which restricts a comprehensive understanding of drug molecules and hampers their resilience against data noise. To overcome the limitations, we construct multimodal deep learning models to cover different molecular representations. We convert drug molecules into three molecular representations, SMILES-encoded vectors, ECFP fingerprints, and molecular graphs. To process the modal information, Transformer-Encoder, bi-directional gated recurrent units (BiGRU), and graph convolutional network (GCN) are utilized for feature learning respectively, which can enhance the model capability to acquire complementary and naturally occurring bioinformatics information. We evaluated our triple-modal model on six molecule datasets. Different from bi-modal learning models, we adopt five fusion methods to capture the specific features and leverage the contribution of each modal information better. Compared with mono-modal models, our multimodal fused deep learning (MMFDL) models outperform single models in accuracy, reliability, and resistance capability against noise. Moreover, we demonstrate its generalization ability in the prediction of binding constants for protein-ligand complex molecules in the refined set of PDBbind. The advantage of the multimodal model lies in its ability to process diverse sources of data using proper models and suitable fusion methods, which would enhance the noise resistance of the model while obtaining data diversity.

Federated learning has emerged as a promising distributed learning paradigm that facilitates collaborative learning among multiple parties without transferring raw data. However, most existing federated learning studies focus on either horizontal or vertical data settings, where the data of different parties are assumed to be from the same feature or sample space. In practice, a common scenario is the hybrid data setting, where data from different parties may differ both in the features and samples. To address this, we propose HybridTree, a novel federated learning approach that enables federated tree learning on hybrid data. We observe the existence of consistent split rules in trees. With the help of these split rules, we theoretically show that the knowledge of parties can be incorporated into the lower layers of a tree. Based on our theoretical analysis, we propose a layer-level solution that does not need frequent communication traffic to train a tree. Our experiments demonstrate that HybridTree can achieve comparable accuracy to the centralized setting with low computational and communication overhead. HybridTree can achieve up to 8 times speedup compared with the other baselines.

We study how to perform unlearning, i.e. forgetting undesirable misbehaviors, on large language models (LLMs). We show at least three scenarios of aligning LLMs with human preferences can benefit from unlearning: (1) removing harmful responses, (2) erasing copyright-protected content as requested, and (3) reducing hallucinations. Unlearning, as an alignment technique, has three advantages. (1) It only requires negative (e.g. harmful) examples, which are much easier and cheaper to collect (e.g. via red teaming or user reporting) than positive (e.g. helpful and often human-written) examples required in RLHF (RL from human feedback). (2) It is computationally efficient. (3) It is especially effective when we know which training samples cause the misbehavior. To the best of our knowledge, our work is among the first to explore LLM unlearning. We are also among the first to formulate the settings, goals, and evaluations in LLM unlearning. We show that if practitioners only have limited resources, and therefore the priority is to stop generating undesirable outputs rather than to try to generate desirable outputs, unlearning is particularly appealing. Despite only having negative samples, our ablation study shows that unlearning can still achieve better alignment performance than RLHF with just 2% of its computational time.

Blockchain smart contracts have given rise to a variety of interesting and compelling applications and emerged as a revolutionary force for the Internet. Quite a few practitioners have devoted themselves to developing tools for detecting bugs in smart contracts. One line of efforts revolve around static analysis techniques, which heavily suffer from high false-positive rates. Another line of works concentrate on fuzzing techniques. Unfortunately, current fuzzing approaches for smart contracts tend to conduct fuzzing starting from the initial state of the contract, which expends too much energy revolving around the initial state and thus is usually unable to unearth bugs triggered by other states. Moreover, most existing methods treat each branch equally, failing to take care of the branches that are rare or more likely to possess bugs. This might lead to resources wasted on normal branches. In this paper, we try to tackle these challenges from three aspects: (1) In generating function invocation sequences, we explicitly consider data dependencies between functions to facilitate exploring richer states. We further prolong a function invocation sequence S1 by appending a new sequence S2, so that S2 can start fuzzing from states that are different from the initial state. (2) We incorporate a branch distance-based measure to evolve test cases iteratively towards a target branch. (3) We engage a branch search algorithm to discover rare and vulnerable branches, and design an energy allocation mechanism to take care of exercising these crucial branches. We implement IR-Fuzz and extensively evaluate it over 12K real-world contracts. Empirical results show that: (i) IR-Fuzz achieves 28% higher branch coverage than state-of-the-art fuzzing approaches, and (ii) IR-Fuzz detects more vulnerabilities and increases the average accuracy of vulnerability detection by 7% over current methods., Comment: This paper has been accepted by IEEE Transactions On Information Forensics And Security (TIFS 2022)

Graph Neural Networks (GNNs) have been widely applied to different tasks such as bioinformatics, drug design, and social networks. However, recent studies have shown that GNNs are vulnerable to adversarial attacks which aim to mislead the node or subgraph classification prediction by adding subtle perturbations. Detecting these attacks is challenging due to the small magnitude of perturbation and the discrete nature of graph data. In this paper, we propose a general adversarial edge detection pipeline EDoG without requiring knowledge of the attack strategies based on graph generation. Specifically, we propose a novel graph generation approach combined with link prediction to detect suspicious adversarial edges. To effectively train the graph generative model, we sample several sub-graphs from the given graph data. We show that since the number of adversarial edges is usually low in practice, with low probability the sampled sub-graphs will contain adversarial edges based on the union bound. In addition, considering the strong attacks which perturb a large number of edges, we propose a set of novel features to perform outlier detection as the preprocessing for our detection. Extensive experimental results on three real-world graph datasets including a private transaction rule dataset from a major company and two types of synthetic graphs with controlled properties show that EDoG can achieve above 0.8 AUC against four state-of-the-art unseen attack strategies without requiring any knowledge about the attack type; and around 0.85 with knowledge of the attack type. EDoG significantly outperforms traditional malicious edge detection baselines. We also show that an adaptive attack with full knowledge of our detection pipeline is difficult to bypass it., Comment: Accepted by IEEE Conference on Secure and Trustworthy Machine Learning 2023

Recent studies show that training deep neural networks (DNNs) with Lipschitz constraints are able to enhance adversarial robustness and other model properties such as stability. In this paper, we propose a layer-wise orthogonal training method (LOT) to effectively train 1-Lipschitz convolution layers via parametrizing an orthogonal matrix with an unconstrained matrix. We then efficiently compute the inverse square root of a convolution kernel by transforming the input domain to the Fourier frequency domain. On the other hand, as existing works show that semi-supervised training helps improve empirical robustness, we aim to bridge the gap and prove that semi-supervised learning also improves the certified robustness of Lipschitz-bounded models. We conduct comprehensive evaluations for LOT under different settings. We show that LOT significantly outperforms baselines regarding deterministic l2 certified robustness, and scales to deeper neural networks. Under the supervised scenario, we improve the state-of-the-art certified robustness for all architectures (e.g. from 59.04% to 63.50% on CIFAR-10 and from 32.57% to 34.59% on CIFAR-100 at radius rho = 36/255 for 40-layer networks). With semi-supervised learning over unlabelled data, we are able to improve state-of-the-art certified robustness on CIFAR-10 at rho = 108/255 from 36.04% to 42.39%. In addition, LOT consistently outperforms baselines on different model architectures with only 1/3 evaluation time., Comment: NeurIPS 2022

Federated Learning (FL) has become a practical and widely adopted distributed learning paradigm. However, the lack of a comprehensive and standardized solution covering diverse use cases makes it challenging to use in practice. In addition, selecting an appropriate FL framework for a specific use case can be a daunting task. In this work, we present UniFed, the first unified platform for standardizing existing open-source FL frameworks. The platform streamlines the end-to-end workflow for distributed experimentation and deployment, encompassing 11 popular open-source FL frameworks. In particular, to address the substantial variations in workflows and data formats, UniFed introduces a configuration-based schema-enforced task specification, offering 20 editable fields. UniFed also provides functionalities such as distributed execution management, logging, and data analysis. With UniFed, we evaluate and compare 11 popular FL frameworks from the perspectives of functionality, privacy protection, and performance, through conducting developer surveys and code-level investigation. We collect 15 diverse FL scenario setups (e.g., horizontal and vertical settings) for FL framework evaluation. This comprehensive evaluation allows us to analyze both model and system performance, providing detailed comparisons and offering recommendations for framework selection. UniFed simplifies the process of selecting and utilizing the appropriate FL framework for specific use cases, while enabling standardized distributed experimentation and deployment. Our results and analysis based on experiments with up to 178 distributed nodes provide valuable system design and deployment insights, aiming to empower practitioners in their pursuit of effective FL solutions., Comment: Code: https://github.com/AI-secure/FLBenchmark-toolkit Website: https://unifedbenchmark.github.io/

Machine learning (ML) robustness and domain generalization are fundamentally correlated: they essentially concern data distribution shifts under adversarial and natural settings, respectively. On one hand, recent studies show that more robust (adversarially trained) models are more generalizable. On the other hand, there is a lack of theoretical understanding of their fundamental connections. In this paper, we explore the relationship between regularization and domain transferability considering different factors such as norm regularization and data augmentations (DA). We propose a general theoretical framework proving that factors involving the model function class regularization are sufficient conditions for relative domain transferability. Our analysis implies that ``robustness" is neither necessary nor sufficient for transferability; rather, regularization is a more fundamental perspective for understanding domain transferability. We then discuss popular DA protocols (including adversarial training) and show when they can be viewed as the function class regularization under certain conditions and therefore improve generalization. We conduct extensive experiments to verify our theoretical findings and show several counterexamples where robustness and generalization are negatively correlated on different datasets., Comment: ICML2022