Semantic loss guided data efficient supervised fine tuning for Safe Responses in LLMs

Large Language Models (LLMs) generating unsafe responses to toxic prompts is a significant issue in their applications. While various efforts aim to address this safety concern, previous approaches often demand substantial human data collection or rely on the less dependable option of using another LLM to generate corrective data. In this paper, we aim to take this problem and overcome limitations of requiring significant high-quality human data. Our method requires only a small set of unsafe responses to toxic prompts, easily obtained from the unsafe LLM itself. By employing a semantic cost combined with a negative Earth Mover Distance (EMD) loss, we guide the LLM away from generating unsafe responses. Additionally, we propose a novel lower bound for EMD loss, enabling more efficient optimization. Our results demonstrate superior performance and data efficiency compared to baselines, and we further examine the nuanced effects of over-alignment and potential degradation of language capabilities when using contrastive data.

Heterogeneous Graph Generation: A Hierarchical Approach using Node Feature Pooling

Heterogeneous graphs are present in various domains, such as social networks, recommendation systems, and biological networks. Unlike homogeneous graphs, heterogeneous graphs consist of multiple types of nodes and edges, each representing different entities and relationships. Generating realistic heterogeneous graphs that capture the complex interactions among diverse entities is a difficult task due to several reasons. The generator has to model both the node type distribution along with the feature distribution for each node type. In this paper, we look into solving challenges in heterogeneous graph generation, by employing a two phase hierarchical structure, wherein the first phase creates a skeleton graph with node types using a prior diffusion based model and in the second phase, we use an encoder and a sampler structure as generator to assign node type specific features to the nodes. A discriminator is used to guide training of the generator and feature vectors are sampled from a node feature pool. We conduct extensive experiments with subsets of IMDB and DBLP datasets to show the effectiveness of our method and also the need for various architecture components.

Towards Neural Network based Cognitive Models of Dynamic Decision-Making by Humans

Modelling human cognitive processes in dynamic decision-making tasks has been an endeavor in AI for a long time. Some initial works have attempted to utilize neural networks (and large language models) but often assume one common model for all humans and aim to emulate human behavior in aggregate. However, behavior of each human is distinct, heterogeneous and relies on specific past experiences in specific tasks. To that end, we build on a well known model of cognition, namely Instance Based Learning (IBL), that posits that decisions are made based on similar situations encountered in the past. We propose two new attention based neural network models to model human decision-making in dynamic settings. We experiment with two distinct datasets gathered from human subject experiment data, one focusing on detection of phishing email by humans and another where humans act as attackers in a cybersecurity setting and decide on an attack option. We conduct extensive experiments with our two neural network models, IBL, and GPT3.5, and demonstrate that one of our neural network models achieves the best performance in representing human decision-making. We find an interesting trend that all models predict a human's decision better if that human is better at the task. We also explore explanation of human decisions based on what our model considers important in prediction. Overall, our work yields promising results for further use of neural networks in cognitive modelling of human decision making. Our code is available at https://github.com/shshnkreddy/NCM-HDM.

Bootstrapping Language Models with DPO Implicit Rewards

Human alignment in large language models (LLMs) is an active area of research. A recent groundbreaking work, direct preference optimization (DPO), has greatly simplified the process from past work in reinforcement learning from human feedback (RLHF) by bypassing the reward learning stage in RLHF. DPO, after training, provides an implicit reward model. In this work, we make a novel observation that this implicit reward model can by itself be used in a bootstrapping fashion to further align the LLM. Our approach is to use the rewards from a current LLM model to construct a preference dataset, which is then used in subsequent DPO rounds. We incorporate refinements that debias the length of the responses and improve the quality of the preference dataset to further improve our approach. Our approach, named self-alignment with DPO ImpliCit rEwards (DICE), shows great improvements in alignment and achieves superior performance than Gemini Pro on AlpacaEval 2, reaching 27.55% length-controlled win rate against GPT-4 Turbo, but with only 8B parameters and no external feedback. Our code is available at https://github.com/sail-sg/dice.

Probabilistic Perspectives on Error Minimization in Adversarial Reinforcement Learning

Deep Reinforcement Learning (DRL) policies are critically vulnerable to adversarial noise in observations, posing severe risks in safety-critical scenarios. For example, a self-driving car receiving manipulated sensory inputs about traffic signs could lead to catastrophic outcomes. Existing strategies to fortify RL algorithms against such adversarial perturbations generally fall into two categories: (a) using regularization methods that enhance robustness by incorporating adversarial loss terms into the value objectives, and (b) adopting "maximin" principles, which focus on maximizing the minimum value to ensure robustness. While regularization methods reduce the likelihood of successful attacks, their effectiveness drops significantly if an attack does succeed. On the other hand, maximin objectives, although robust, tend to be overly conservative. To address this challenge, we introduce a novel objective called Adversarial Counterfactual Error (ACoE), which naturally balances optimizing value and robustness against adversarial attacks. To optimize ACoE in a scalable manner in model-free settings, we propose a theoretically justified surrogate objective known as Cumulative-ACoE (C-ACoE). The core idea of optimizing C-ACoE is utilizing the belief about the underlying true state given the adversarially perturbed observation. Our empirical evaluations demonstrate that our method outperforms current state-of-the-art approaches for addressing adversarial RL problems across all established benchmarks (MuJoCo, Atari, and Highway) used in the literature.

Generative Modelling of Stochastic Actions with Arbitrary Constraints in Reinforcement Learning

Many problems in Reinforcement Learning (RL) seek an optimal policy with large discrete multidimensional yet unordered action spaces; these include problems in randomized allocation of resources such as placements of multiple security resources and emergency response units, etc. A challenge in this setting is that the underlying action space is categorical (discrete and unordered) and large, for which existing RL methods do not perform well. Moreover, these problems require validity of the realized action (allocation); this validity constraint is often difficult to express compactly in a closed mathematical form. The allocation nature of the problem also prefers stochastic optimal policies, if one exists. In this work, we address these challenges by (1) applying a (state) conditional normalizing flow to compactly represent the stochastic policy -- the compactness arises due to the network only producing one sampled action and the corresponding log probability of the action, which is then used by an actor-critic method; and (2) employing an invalid action rejection method (via a valid action oracle) to update the base policy. The action rejection is enabled by a modified policy gradient that we derive. Finally, we conduct extensive experiments to show the scalability of our approach compared to prior methods and the ability to enforce arbitrary state-conditional constraints on the support of the distribution of actions in any state.

Conditioning Hierarchical Reinforcement Learning on Flexible Constraints

Safety in goal directed Reinforcement Learning (RL) settings has typically been handled through constraints over trajectories and have demonstrated good performance in primarily short horizon tasks (goal is not too far away). In this paper, we are specifically interested in the problem of solving temporally extended decision making problems such as (1) robots that have to clean different areas in a house while avoiding slippery and unsafe areas (e.g., stairs) and retaining enough charge to move to a charging dock; (2) autonomous electric vehicles that have to reach a far away destination while having to optimize charging locations along the way; in the presence of complex safety constraints. Our key contribution is a (safety) Constrained Planning with Reinforcement Learning (CoP-RL) mechanism that combines a high-level constrained planning agent (which computes a reward maximizing path from a given start to a far away goal state while satisfying cost constraints) with a low-level goal conditioned RL agent (which estimates cost and reward values to move between nearby states). A major advantage of CoP-RL is that it can handle constraints on the cost value distribution (e.g., on Conditional Value at Risk, CVaR, and also on expected value). We perform extensive experiments with different types of safety constraints to demonstrate the utility of our approach over leading best approaches in constrained and hierarchical RL.

Mind Your Data! Hiding Backdoors in Offline Reinforcement Learning Datasets

A growing body of research works has focused on the Offline Reinforcement Learning (RL) paradigm. Data providers share large pre-collected datasets on which others can train high-quality agents without interacting with the environments. Such an offline RL paradigm has demonstrated effectiveness in many critical tasks, including robot control, autonomous driving, etc. A well-trained agent can be regarded as a software system. However, less attention is paid to investigating the security threats to the offline RL system. In this paper, we focus on a critical security threat: backdoor attacks. Given normal observations, an agent implanted with backdoors takes actions leading to high rewards. However, the same agent takes actions that lead to low rewards if the observations are injected with triggers that can activate the backdoor. In this paper, we propose Baffle (Backdoor Attack for Offline Reinforcement Learning) and evaluate how different Offline RL algorithms react to this attack. Our experiments conducted on four tasks and four offline RL algorithms expose a disquieting fact: none of the existing offline RL algorithms is immune to such a backdoor attack. More specifically, Baffle modifies $10\%$ of the datasets for four tasks (3 robotic controls and 1 autonomous driving). Agents trained on the poisoned datasets perform well in normal settings. However, when triggers are presented, the agents' performance decreases drastically by $63.6\%$, $57.8\%$, $60.8\%$ and $44.7\%$ in the four tasks on average. The backdoor still persists after fine-tuning poisoned agents on clean datasets. We further show that the inserted backdoor is also hard to be detected by a popular defensive method. This paper calls attention to developing more effective protection for the open-source offline RL dataset.

Scalable Distributional Robustness in a Class of Non Convex Optimization with Guarantees

Distributionally robust optimization (DRO) has shown lot of promise in providing robustness in learning as well as sample based optimization problems. We endeavor to provide DRO solutions for a class of sum of fractionals, non-convex optimization which is used for decision making in prominent areas such as facility location and security games. In contrast to previous work, we find it more tractable to optimize the equivalent variance regularized form of DRO rather than the minimax form. We transform the variance regularized form to a mixed-integer second order cone program (MISOCP), which, while guaranteeing near global optimality, does not scale enough to solve problems with real world data-sets. We further propose two abstraction approaches based on clustering and stratified sampling to increase scalability, which we then use for real world data-sets. Importantly, we provide near global optimality guarantees for our approach and show experimentally that our solution quality is better than the locally optimal ones achieved by state-of-the-art gradient-based methods. We experimentally compare our different approaches and baselines, and reveal nuanced properties of a DRO solution.

Proceedings of the Artificial Intelligence for Cyber Security (AICS) Workshop at AAAI 2022

The workshop will focus on the application of AI to problems in cyber security. Cyber systems generate large volumes of data, utilizing this effectively is beyond human capabilities. Additionally, adversaries continue to develop new attacks. Hence, AI methods are required to understand and protect the cyber domain. These challenges are widely studied in enterprise networks, but there are many gaps in research and practice as well as novel problems in other domains. In general, AI techniques are still not widely adopted in the real world. Reasons include: (1) a lack of certification of AI for security, (2) a lack of formal study of the implications of practical constraints (e.g., power, memory, storage) for AI systems in the cyber domain, (3) known vulnerabilities such as evasion, poisoning attacks, (4) lack of meaningful explanations for security analysts, and (5) lack of analyst trust in AI solutions. There is a need for the research community to develop novel solutions for these practical issues.