Red-teaming, or identifying prompts that elicit harmful responses, is a critical step in ensuring the safe and responsible deployment of large language models (LLMs). Developing effective protection against many modes of attack prompts requires discovering diverse attacks. Automated red-teaming typically uses reinforcement learning to fine-tune an attacker language model to generate prompts that elicit undesirable responses from a target LLM, as measured, for example, by an auxiliary toxicity classifier. We show that even with explicit regularization to favor novelty and diversity, existing approaches suffer from mode collapse or fail to generate effective attacks. As a flexible and probabilistically principled alternative, we propose to use GFlowNet fine-tuning, followed by a secondary smoothing phase, to train the attacker model to generate diverse and effective attack prompts. We find that the attacks generated by our method are effective against a wide range of target LLMs, both with and without safety tuning, and transfer well between target LLMs. Finally, we demonstrate that models safety-tuned using a dataset of red-teaming prompts generated by our method are robust to attacks from other RL-based red-teaming approaches.
相關內容
Error control by means of a posteriori error estimators or indica-tors and adaptive discretizations, such as adaptive mesh refinement, have emerged in the late seventies. Since then, numerous theoretical developments and improvements have been made, as well as the first attempts to introduce them into real-life industrial applications. The present introductory chapter provides an overview of the subject, highlights some of the achievements to date and discusses possible perspectives.
The spatial auditory attention decoding (Sp-AAD) technology aims to determine the direction of auditory attention in multi-talker scenarios via neural recordings. Despite the success of recent Sp-AAD algorithms, their performance is hindered by trial-specific features in EEG data. This study aims to improve decoding performance against these features. Studies in neuroscience indicate that spatial auditory attention can be reflected in the topological distribution of EEG energy across different frequency bands. This insight motivates us to propose Prototype Training, a neuroscience-inspired method for Sp-AAD. This method constructs prototypes with enhanced energy distribution representations and reduced trial-specific characteristics, enabling the model to better capture auditory attention features. To implement prototype training, an EEGWaveNet that employs the wavelet transform of EEG is further proposed. Detailed experiments indicate that the EEGWaveNet with prototype training outperforms other competitive models on various datasets, and the effectiveness of the proposed method is also validated. As a training method independent of model architecture, prototype training offers new insights into the field of Sp-AAD.
Recent advances in active materials and fabrication techniques have enabled the production of cyclically self-deployable metamaterials with an expanded functionality space. However, designing metamaterials that possess continuously tunable mechanical properties after self-deployment remains a challenge, notwithstanding its importance. Inspired by push puppets, we introduce an efficient design strategy to create reversibly self-deployable metamaterials with continuously tunable post-deployment stiffness and damping. Our metamaterial comprises contracting actuators threaded through beads with matching conical concavo-convex interfaces in networked chains. The slack network conforms to arbitrary shapes, but when actuated, it self-assembles into a preprogrammed configuration with beads gathered together. Further contraction of the actuators can dynamically tune the assembly's mechanical properties through the beads' particle jamming, while maintaining the overall structure with minimal change. We show that, after deployment, such metamaterials exhibit pronounced tunability in bending-dominated configurations: they can become more than 35 times stiffer and change their damping capability by over 50%. Through systematic analysis, we find that the beads'conical angle can introduce geometric nonlinearity, which has a major effect on the self-deployability and tunability of the metamaterial. Our work provides routes towards reversibly self-deployable, lightweight, and tunable metamaterials, with potential applications in soft robotics, reconfigurable architectures, and space engineering.
Urban congestion remains a critical challenge, with traffic signal control (TSC) emerging as a potent solution. TSC is often modeled as a Markov Decision Process problem and then solved using reinforcement learning (RL), which has proven effective. However, the existing RL-based TSC system often overlooks imperfect observations caused by degraded communication, such as packet loss, delays, and noise, as well as rare real-life events not included in the reward function, such as unconsidered emergency vehicles. To address these limitations, we introduce a novel integration framework that combines a large language model (LLM) with RL. This framework is designed to manage overlooked elements in the reward function and gaps in state information, thereby enhancing the policies of RL agents. In our approach, RL initially makes decisions based on observed data. Subsequently, LLMs evaluate these decisions to verify their reasonableness. If a decision is found to be unreasonable, it is adjusted accordingly. Additionally, this integration approach can be seamlessly integrated with existing RL-based TSC systems without necessitating modifications. Extensive testing confirms that our approach reduces the average waiting time by $17.5\%$ in degraded communication conditions as compared to traditional RL methods, underscoring its potential to advance practical RL applications in intelligent transportation systems. The related code can be found at \url{//github.com/Traffic-Alpha/iLLM-TSC}.
Large language models (LLMs) demonstrate general intelligence across a variety of machine learning tasks, thereby enhancing the commercial value of their intellectual property (IP). To protect this IP, model owners typically allow user access only in a black-box manner, however, adversaries can still utilize model extraction attacks to steal the model intelligence encoded in model generation. Watermarking technology offers a promising solution for defending against such attacks by embedding unique identifiers into the model-generated content. However, existing watermarking methods often compromise the quality of generated content due to heuristic alterations and lack robust mechanisms to counteract adversarial strategies, thus limiting their practicality in real-world scenarios. In this paper, we introduce an adaptive and robust watermarking method (named ModelShield) to protect the IP of LLMs. Our method incorporates a self-watermarking mechanism that allows LLMs to autonomously insert watermarks into their generated content to avoid the degradation of model content. We also propose a robust watermark detection mechanism capable of effectively identifying watermark signals under the interference of varying adversarial strategies. Besides, ModelShield is a plug-and-play method that does not require additional model training, enhancing its applicability in LLM deployments. Extensive evaluations on two real-world datasets and three LLMs demonstrate that our method surpasses existing methods in terms of defense effectiveness and robustness while significantly reducing the degradation of watermarking on the model-generated content.
Machine unlearning techniques, which involve retracting data records and reducing influence of said data on trained models, help with the user privacy protection objective but incur significant computational costs. Weight perturbation-based unlearning is a general approach, but it typically involves globally modifying the parameters. We propose fine-grained Top-K and Random-k parameters perturbed inexact machine unlearning strategies that address the privacy needs while keeping the computational costs tractable. In order to demonstrate the efficacy of our strategies we also tackle the challenge of evaluating the effectiveness of machine unlearning by considering the model's generalization performance across both unlearning and remaining data. To better assess the unlearning effect and model generalization, we propose novel metrics, namely, the forgetting rate and memory retention rate. However, for inexact machine unlearning, current metrics are inadequate in quantifying the degree of forgetting that occurs after unlearning strategies are applied. To address this, we introduce SPD-GAN, which subtly perturbs the distribution of data targeted for unlearning. Then, we evaluate the degree of unlearning by measuring the performance difference of the models on the perturbed unlearning data before and after the unlearning process. By implementing these innovative techniques and metrics, we achieve computationally efficacious privacy protection in machine learning applications without significant sacrifice of model performance. Furthermore, this approach provides a novel method for evaluating the degree of unlearning.
We discuss a connection between a generative model, called the diffusion model, and nonequilibrium thermodynamics for the Fokker-Planck equation, called stochastic thermodynamics. Based on the techniques of stochastic thermodynamics, we derive the speed-accuracy trade-off for the diffusion models, which is a trade-off relationship between the speed and accuracy of data generation in diffusion models. Our result implies that the entropy production rate in the forward process affects the errors in data generation. From a stochastic thermodynamic perspective, our results provide quantitative insight into how best to generate data in diffusion models. The optimal learning protocol is introduced by the conservative force in stochastic thermodynamics and the geodesic of space by the 2-Wasserstein distance in optimal transport theory. We numerically illustrate the validity of the speed-accuracy trade-off for the diffusion models with different noise schedules such as the cosine schedule, the conditional optimal transport, and the optimal transport.
We discuss a connection between a generative model, called the diffusion model, and nonequilibrium thermodynamics for the Fokker-Planck equation, called stochastic thermodynamics. Based on the techniques of stochastic thermodynamics, we derive the speed-accuracy trade-off for the diffusion models, which is a trade-off relationship between the speed and accuracy of data generation in diffusion models. Our result implies that the entropy production rate in the forward process affects the errors in data generation. From a stochastic thermodynamic perspective, our results provide quantitative insight into how best to generate data in diffusion models. The optimal learning protocol is introduced by the conservative force in stochastic thermodynamics and the geodesic of space by the 2-Wasserstein distance in optimal transport theory. We numerically illustrate the validity of the speed-accuracy trade-off for the diffusion models with different noise schedules such as the cosine schedule, the conditional optimal transport, and the optimal transport.
This research presents a comprehensive approach to predicting the duration of traffic incidents and classifying them as short-term or long-term across the Sydney Metropolitan Area. Leveraging a dataset that encompasses detailed records of traffic incidents, road network characteristics, and socio-economic indicators, we train and evaluate a variety of advanced machine learning models including Gradient Boosted Decision Trees (GBDT), Random Forest, LightGBM, and XGBoost. The models are assessed using Root Mean Square Error (RMSE) for regression tasks and F1 score for classification tasks. Our experimental results demonstrate that XGBoost and LightGBM outperform conventional models with XGBoost achieving the lowest RMSE of 33.7 for predicting incident duration and highest classification F1 score of 0.62 for a 30-minute duration threshold. For classification, the 30-minute threshold balances performance with 70.84% short-term duration classification accuracy and 62.72% long-term duration classification accuracy. Feature importance analysis, employing both tree split counts and SHAP values, identifies the number of affected lanes, traffic volume, and types of primary and secondary vehicles as the most influential features. The proposed methodology not only achieves high predictive accuracy but also provides stakeholders with vital insights into factors contributing to incident durations. These insights enable more informed decision-making for traffic management and response strategies. The code is available by the link: //github.com/Future-Mobility-Lab/SydneyIncidents
Conventional entity typing approaches are based on independent classification paradigms, which make them difficult to recognize inter-dependent, long-tailed and fine-grained entity types. In this paper, we argue that the implicitly entailed extrinsic and intrinsic dependencies between labels can provide critical knowledge to tackle the above challenges. To this end, we propose \emph{Label Reasoning Network(LRN)}, which sequentially reasons fine-grained entity labels by discovering and exploiting label dependencies knowledge entailed in the data. Specifically, LRN utilizes an auto-regressive network to conduct deductive reasoning and a bipartite attribute graph to conduct inductive reasoning between labels, which can effectively model, learn and reason complex label dependencies in a sequence-to-set, end-to-end manner. Experiments show that LRN achieves the state-of-the-art performance on standard ultra fine-grained entity typing benchmarks, and can also resolve the long tail label problem effectively.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191