New Zealand


As reinforcement learning (RL) systems are deployed in various safety-critical applications, it is imperative to understand how vulnerable they are to adversarial attacks. Of these, an environment-poisoning attack (EPA) is considered particularly insidious, since environment hyper-parameters are significant factors in determining an RL policy, yet prone to be accessed by third parties. The success of existing EPAs relies on comprehensive prior knowledge of the attacked RL system, including RL agent’s learning mechanism and/or its environment model. Unfortunately, such an assumption of prior knowledge creates an unrealistic attack, one that poses little threat to real-world RL systems.

In this paper, we propose a Double-Black-Box EPA framework, only assuming the attacker&#39;s ability to alter environment hyper-parameters. Considering that environment alteration comes at a cost, we seek minimal poisoning in an unknown environment, that aims to force a black-box RL agent to learn an attacker-designed policy. To this end, we incorporate an inference module in our framework to capture the internal information of the unknown RL system and, accordingly, learn an adaptive strategy based on an approximation of our attack objective. We empirically show the threat posed by our attack to both tabular-RL and deep-RL algorithms, in both discrete and continuous environments.

AAMAS 2022

Spiking Pitch Black: Poisoning an Unknown Environment to Attack Unknown Reinforcement Learners

environment poisoning

security

reinforcement learning

 **RTSS 2021** has successfully concluded. Hope you have enjoyed it, and look forward to seeing you next year in person.
    Announcement for the Awards:
 * Best Paper Award goes to: Monte Carlo Response-Time Analysis, Sergey Bozhko, Georg von der Brüggen, and Björn Brandenburg
 * Best Student Paper Award goes to: Automatic Energy-Hotspot Detection and Elimination in Real-Time Deeply Embedded Systems, Mohsen Shekarisaz, Lothar Thiele, and Mehdi Kargahi.
 * Best Solution Award for Industry Session goes to: Toward Real-Time Guaranteed Scheduling for Autonomous Driving Systems. Jinghao Sun, Tianyi Wang, Kailu Duan, Bin Lu, Jiankang Ren, Zhishan Guo, and Guozhen Tan.



IEEE RTSS 2021

The IEEE Real-Time Systems Symposium (RTSS) is the premier conference in the field of real-time systems, and is a venue for researchers and practitioners to showcase innovations covering all aspects of real-time systems, including theory, design, analysis, implementation, evaluation, and experience. 


As reinforcement learning (RL) systems are deployed in various safety-critical applications, it is imperative to understand how vulnerable they are to adversarial attacks. Of these, an environment-poisoning attack (EPA) is considered particularly insidious, since environment hyper-parameters are significant factors in determining an RL policy, yet prone to be accessed by third parties. The success of existing EPAs relies on comprehensive prior knowledge of the attacked RL system, including RL agent’s learning mechanism and/or its environment model. Unfortunately, such an assumption of prior knowledge creates an unrealistic attack, one that poses little threat to real-world RL systems.

In this paper, we propose a Double-Black-Box EPA framework, only assuming the attacker's ability to alter environment hyper-parameters. Considering that environment alteration comes at a cost, we seek minimal poisoning in an unknown environment, that aims to force a black-box RL agent to learn an attacker-designed policy. To this end, we incorporate an inference module in our framework to capture the internal information of the unknown RL system and, accordingly, learn an adaptive strategy based on an approximation of our attack objective. We empirically show the threat posed by our attack to both tabular-RL and deep-RL algorithms, in both discrete and continuous environments.

technical paper

AAMAS (International Conference on Autonomous Agents and Multiagent Systems) is the largest and most influential conference in the area of agents and multiagent systems. The aim of the conference is to bring together researchers and practitioners in all areas of agent technology and to provide a single, high-profile, internationally renowned forum for research in the theory and practice of autonomous agents and multiagent systems. AAMAS is the flagship conference of the non-profit International Foundation for Autonomous Agents and Multiagent Systems (IFAAMAS).

The AAMAS conference series was initiated in 2002 in Bologna, Italy as a joint event comprising the 6th International Conference on Autonomous Agents (AA), the 5th International Conference on Multiagent Systems (ICMAS), and the 9th International Workshop on Agent Theories, Architectures, and Languages (ATAL).

Subsequent AAMAS conferences have been held in Melbourne, Australia (July 2003), New York City, NY, USA (July 2004), Utrecht, The Netherlands (July 2005), Hakodate, Japan (May 2006), Honolulu, Hawaii, USA (May 2007), Estoril, Portugal (May 2008), Budapest, Hungary (May 2009), Toronto, Canada (May 2010), Taipei, Taiwan (May 2011), Valencia, Spain (June 2012), Minnesota, USA (May 2013), Paris, France (May 2014), Istanbul, Turkey (May 2015), Singapore (May 2016), São Paulo (2017) and Stockholm, Sweden (2018), Montreal (May 2019), Auckland (May 2020, Virtual), London (May 2021, Virtual).
<br>
<br>



A ticket is required to attend this event, please register using the link below:

https://aamas2022-conference.auckland.ac.nz/attending/registration/

Registration Is Required

AAMAS (International Conference on Autonomous Agents and Multiagent Systems) is the largest and most influential conference in the area of agents and multiagent systems. The aim of the conference is to bring together researchers and practitioners in all areas of agent technology and to provide a single, high-profile, internationally renowned forum for research in the theory and practice of autonomous agents and multiagent systems. AAMAS is the flagship conference of the non-profit International Foundation for Autonomous Agents and Multiagent Systems (IFAAMAS).


Although learning has found wide application in multi-agent systems, its effects on the temporal evolution of a system are far from understood. This paper focuses on the dynamics of Q-learning in large-scale multi-agent systems modeled as population games. We revisit the replicator equation model for Q-learning dynamics and observe that this model is inappropriate for our concerned setting. Motivated by this, we develop a new formal model, which bears a formal connection with the continuity equation in physics. 
We show that our model always accurately describes the Q-learning dynamics in population games across different initial settings of MASs and game configurations. We also show that our model can be applied to different exploration mechanisms, describe the mean dynamics, and be extended to Q-learning in 2-player and n-player games. Last but not least, we show that our model can be used to gain insights into algorithm parameters, which potentially can facilitate parameter tuning.

The Dynamics of Q-learning in Population Games: a Physics-inspired Continuity Equation Model


A key challenge of evolutionary game theory and multi-agent learning is to characterize the limit behavior of game dynamics. Whereas convergence is often a property of learning algorithms in games satisfying a particular reward structure (e.g. zero-sum), it is well known, that for general payoffs even basic learning models, such as the replicator dynamics, are not guaranteed to converge. Worse yet, chaotic behavior is possible even in rather simple games, such as variants of Rock-Paper-Scissors games (Sato et al. 2002). Although chaotic behavior in learning dynamics can be precluded by the celebrated Poincaré-Bendixson theorem, it is only applicable to low-dimensional settings. Are there other characteristics of a game, which can force regularity in the limit sets of learning?

We show that behaviors consistent with the Poincaré-Bendixson theorem (limit cycles, but no chaotic attractor) follows purely based on the topological structure of the interaction graph, even for high-dimensional settings with arbitrary number of players and arbitrary payoff matrices. We prove our result for a wide class of follow-the-regularized leader (FoReL) dynamics, which generalize replicator dynamics, for games where each player has two strategies at disposal,and for interaction graphs where payoffs of each agent are only affected by one other agent. Since chaos has been observed in a game with only two players and three strategies, this class of non-chaotic games is in a sense maximal. Moreover, we provide simple conditions, under which such behavior translates to efficiency guarantees, implying that FoReL learning achieves time average sum of payoffs at least as good as that of a Nash equilibrium; and connecting the topology of the dynamics to the social welfare analysis.

Poincaré-Bendixson Limit Sets in Multi-Agent Learning


Concave Utility Reinforcement Learning (CURL) extends RL from linear to concave utilities in the occupancy measure induced by the agent's policy. This encompasses not only RL but also imitation learning and exploration, among others. Yet, this more general paradigm invalidates the classical Bellman equations, and calls for new algorithms. Mean-field Games (MFGs) are a continuous approximation of many-agent RL. They consider the limit case of a continuous distribution of identical agents, anonymous with symmetric interests, and reduce the problem to the study of a single representative agent in interaction with the full population. Our core contribution consists in showing that CURL is a subclass of MFGs. We think this important to bridge together both communities. It also allows to shed light on aspects of both fields: we show the equivalence between concavity in CURL and monotonicity in the associated MFG, between optimality conditions in CURL and Nash equilibrium in MFG, or that Fictitious Play (FP) for this class of MFGs is simply Frank-Wolfe, bringing the first convergence rate for discrete-time FP for MFGs. We also experimentally demonstrate that, using algorithms recently introduced for solving MFGs, we can address the CURL problem more efficiently.

Concave Utility Reinforcement Learning: the Mean-field Game viewpoint


Recent advances in multiagent learning have seen the introduction of a family of algorithms that revolve around the population-based training method PSRO, showing convergence to Nash, correlated and coarse correlated equilibria. 
Notably, when the number of agents increases, learning best-responses becomes exponentially more difficult, and as such hampers PSRO training methods. 
The field of mean-field games provides an asymptotic solution to this problem when the considered games are anonymous-symmetric.
Unfortunately, the mean-field approximation introduces non-linearities which prevent a straightforward adaptation of PSRO.
Building upon optimization and adversarial regret minimization, this paper sidesteps this issue and introduces mean-field PSRO, an adaptation of PSRO which learns Nash, coarse correlated and correlated equilibria in mean-field games. 
The key is to replace the exact distribution computation step by newly-defined mean-field no-adversarial-regret learners, or by black-box optimization. 
We compare the asymptotic complexity of the approach to standard PSRO, greatly improve empirical bandit convergence speed by compressing temporal mixture weights, and ensure it is theoretically robust to payoff noise.
Finally, we illustrate the speed and accuracy of mean-field PSRO on several mean-field games, demonstrating convergence to strong and weak equilibria.

Learning Equilibria in Mean-Field Games: Introducing Mean-Field PSRO


As one of the solutions to the decentralized partially observable Markov decision process (Dec-POMDP) problems, the value decomposition method has achieved significant results recently. However, most value decomposition methods require the fully observable state of the environment during training, but this is not feasible in some scenarios where only incomplete and noisy observations can be obtained. Therefore, we propose a novel value decomposition framework, named State Inference for value DEcomposition (SIDE), which eliminates the need to know the global state by simultaneously seeking solutions to the two problems of optimal control and state inference. SIDE can be extended to any value decomposition method to tackle partially observable problems. By comparing with the performance of different algorithms in StarCraft II micromanagement tasks, we verified that though without accessible states, SIDE can infer the current state that contributes to the reinforcement learning process based on past local observations and even achieve superior results to many baselines in some complex scenarios.

SIDE: State Inference for Partially Observable Cooperative Multi-Agent Reinforcement Learning


Intrinsic rewards can improve exploration in reinforcement learning, but often suffer from instability caused by non-stationary reward shaping and strong dependency on hyperparameters. In this work, we introduce Decoupled RL (DeRL) as a general framework which trains separate policies for intrinsically motivated exploration and exploitation. Such decoupling allows DeRL to leverage the benefits of intrinsic rewards for exploration while demonstrating improved robustness and sample efficiency. We evaluate DeRL algorithms in two sparse-reward environments with multiple types of intrinsic rewards. Our results show that DeRL is more robust to varying scale and speed of decay of intrinsic rewards and converges to the same evaluation returns than intrinsically motivated baselines in fewer interactions. Lastly, we discuss the challenge of distribution shift and show that divergence constraint regularisers can successfully minimise instability caused by divergence of exploration and exploitation policies.

Decoupled Reinforcement Learning to Stabilise Intrinsically-Motivated Exploration


Multi-agent reinforcement learning holds the key for solving complex tasks that demand coordination of learning agents. However, strong coordination often leads to expensive exploration over the exponentially large state-action space. A powerful approach is to decompose team works into roles, which are ideally assigned to agents with the relevant skills. Training agents to adaptively choose and play emerging roles in team thus allows the team to scale to complex tasks and quickly adapt to changing environments. These promises, however, have not been fully realized by current role-based multi-agent reinforcement learning methods as they assume either a pre-defined role structure or a fixed team size. We propose a framework to learn role assignment and transfer across team sizes. In particular, we train a role assignment network for small teams by demonstration and transfer the network to larger teams, which continue to learn through interaction with the environment. We demonstrate that re-using the role-based credit assignment structure can foster the learning process of larger reinforcement learning teams to achieve tasks requiring different roles. Our proposal outperforms competing techniques in enriched role-enforcing Prey-Predator games and in new scenarios in the StarCraft II Micro-Management benchmark.

Learning to Transfer Role Assignment Across Team Sizes


Identifying datapoints that substantially differ from normality is the task of anomaly detection (AD).
While AD has gained widespread attention in rich data domains such as images, videos, audio and text, it has has been studied less frequently in the context of reinforcement learning (RL). This is due to the additional layer of complexity that RL introduces through sequential decision making. Developing suitable anomaly detectors for RL is of particular importance in safety-critical scenarios where acting on anomalous data could result in hazardous situations. 
In this work, we address the question of what AD means in the context of RL. We found that current research trains and evaluates on overly simplistic and unrealistic scenarios which reduce to classic pattern recognition tasks. We link AD in RL to various fields in RL such as lifelong RL and generalization. We discuss their similarities, differences, and how the fields can benefit from each other. Moreover, we identify non-stationarity to be one of the key drivers for future research on AD in RL and make a first step towards a more formal treatment of the problem by framing it in terms of the recently introduced block contextual Markov decision process.
Finally, define a list of practical desiderata for future problems in the field.

Spiking Pitch Black: Poisoning an Unknown Environment to Attack Unknown Reinforcement Learners

Downloads

Next from AAMAS 2022

The Dynamics of Q-learning in Population Games: a Physics-inspired Continuity Equation Model

Similar lecture

Catch Me If You Learn: Real-Time Attack Detection and Mitigation in Learning Enabled CPS

Spiking Pitch Black: Poisoning an Unknown Environment to Attack Unknown Reinforcement Learners

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from AAMAS 2022

The Dynamics of Q-learning in Population Games: a Physics-inspired Continuity Equation Model

Similar lecture

Catch Me If You Learn: Real-Time Attack Detection and Mitigation in Learning Enabled CPS

Downloads