Trustworthy and Responsible AI

• Mathematical foundations of trustworthy and responsible AI (e.g., learning techniques, causality, information theory)
• Trustworthy and responsible AI metrics and their interconnections
• Theoretical understanding of trustworthy and responsible AI
• Trustworthy and responsible AI in the real world
• Trustworthy and responsible AI engineering practice
• Trustworthy and responsible AI for large foundation models
• Auditability & Accountability
• Trustworthy human-computer interactions
• Social Acceptance of trustworthy and responsible AI
• Fairness and bias
• Risk assessment and risk-aware decision making

Abstract: Cooperative game theorists propose the following attractive process: (1) capture the abstract value} of each possible coalition} of individuals, (2) write down some principles, or axioms, on how to distribute the value (e.g., allocate importance to features or parameters), and then, (3) find a set of allocations that satisfy the principles. The Shapley value has received much attention -- but it is just one solution concept, satisfying one set of principles, in one class of games. It is popular among game theorists because the axioms, and the class of TU-games, are reasonable in game theory. In AI and ML, we should choose carefully what is reasonable for our own purposes.
In this paper, we highlight solution concepts in the class of multi-choice games (MC-games). These are model agnostic, and unique to their own set of axioms, just like the Shapley value. This paper offers a general algorithm for constructing any MC-game framework with polynomial time complexity in the number of parameter levels, and an application of this algorithm that is transparent, and can be readily generalised to local explanation frameworks such as SHapley Additive exPlanations (SHAP).

Authors: Daniel Fryer (The University of Queensland); Hien Nguyen (University of Queensland & School of Mathematics and Physics); David Lowing (Université Paris-Saclay); Inga Strumke (NTNU)

Abstract: Emotion recognition in textual conversation (ERTC) is crucial for developing advanced conversational systems that can understand and support users' emotional needs. Despite significant progress in ERTC using deep learning techniques, the subjective nature of emotion and the lack of emotional cues in textual conversations pose challenges in building highly accurate systems. In this paper, we propose an uncertainty-aware approach to ERTC, employing an approximate Bayesian inference method to address the inherent uncertainty in emotion classification. We provide confidence metrics for individual predictions using conformal prediction and standard error mean (SEM), enabling downstream tasks to make informed decisions based on the level of confidence of each prediction. Our approach aims to enhance the reliability and trustworthiness of ERTC systems, paving the way for their wider adoption in real-world applications.

Authors: Samad Roohi (La Trobe University); Richard Skarbez (La Trobe University); Hien Nguyen (University of Queensland & School of Mathematics and Physics)

Abstract: Recommender systems have shown to popular in many Internet communities, as they could help users discover interesting items based on their history behaviors. However, with the explosive growth of data-intensive tasks and online information, cybersecurity risks become larger, conventional collaborative recommendation algorithms may not meet users’ security requirements. Besides, the sparsity issue and the cold-start issue also hinder the performance of conventional recommendation methods. Recently, deep learning has shown to outperform traditional modelling techniques, which can be employed in Recommender systems (RSs) to improve user behavior prediction. In light of these challenges and observations, an intelligent recommendation method based on multi-interest network and responsible deep learning is proposed. It utilizes multi-source behavior information to improve prediction performance, where multi-view preference embeddings, including self-embeddings, interaction-aware embeddings, and neighbor-based embeddings, are combined to model users’ interests at a finer granularity. Specifically, two factorization techniques, Matrix factorization (MF) and Tensor Factorization (TF), are applied to mine local and global interactions between users and items for self-embedding learning. Moreover, interaction with context and neighbor-based interest are also considered to improve the modeling of user preferences. In neighbor-based embedding learning, a responsible search scheme is adopted to fast similarity searching and support privacy preservation. Finally, a DNN-based prediction mechanism is adopted for embedding aggregation and final prediction. Extensive experiments on real-world datasets show that our proposal achieves decent prediction performance with security concerns compared with state-of the-art baselines.

Authors: Shunmei Meng (Nanjing University of Science and Technology)*; Xiao Liu (Nanjing University of Science and Technology); Xuyun Zhang (Macquarie University)

Abstract: Federated Learning (FL) enables many clients to train a joint model without sharing the raw data. While many byzantine-robust FL methods have been proposed, FL remains vulnerable to security attacks such as poisoning attacks and evasion attacks due to its distributed adversarial environment. Additionally, real-world training data used in FL are usually Non-Independent and Identically Distributed (Non-IID), which further weakens the robustness of the existing FL methods (such as Krum, Median, Trimmed-Mean, etc.), thereby making it possible for a global model in FL to be broken in extreme Non-IID scenarios. In this work, we mitigate the aforementioned weaknesses of existing FL methods in Non-IID and adversarial scenarios by proposing a new FL framework called Mini Federated Learning (Mini-FL). Mini-FL follows the general FL approach but considers the Non-IID sources of FL and aggregates the gradients by groups. Specifically, Mini-FL first performs unsupervised learning for the gradients received to define the grouping policy. Then, the server divides the gradients received into different groups according to the grouping policy defined and performs byzantine-robust aggregation. Finally, the server calculates the weighted mean of gradients from each group to update the global model. Owning the strong generality, Mini-FL can utilize the most existing byzantine-robust method. We demonstrate that Mini-FL effectively enhances FL robustness and achieves greater global accuracy than existing FL methods when against security attacks and in Non-IID settings

Authors: Yanli Li (The University of Sydney)