AAAI 2022 Papers

We report a new neural backdoor attack, named Hibernated Backdoor, which is stealthy, aggressive and devastating. The backdoor is planted in a hibernated mode to avoid being detected. Once deployed and fine-tuned on end-devices, the hibernated backdoor turns into the active state that can be exploited by the attacker. To the best of our knowledge, this is the first hibernated neural backdoor attack. It is achieved by maximizing the mutual information (MI) between the gradients of regular and malicious data on the model. We introduce a practical algorithm to achieve MI maximization to effectively plant the hibernated backdoor. To evade adaptive defenses, we further develop a targeted hibernated backdoor, which can only be activated by specific data samples and thus achieves a higher degree of stealthiness. We show the hibernated backdoor is robust and cannot be removed by existing backdoor removal schemes. It has been fully tested on four datasets with two neural network architectures, compared to five existing backdoor attacks, and evaluated using seven backdoor detection schemes. The experiments demonstrate the effectiveness of the hibernated backdoor attack under various settings.

Human conversations are guided by short-term and long-term goals. We study how to plan short-term goal sequences as coherently as humans do and naturally direct them to an assigned long-term goal in open-domain conversations. Goal sequences are a series of knowledge graph (KG) entity-relation connections generated by KG walkers that traverse through the KG. The existing recurrent and graph attention based KG walkers either insufficiently utilize the conversation states or lack global guidance. In our work, a hierarchical model learns goal planning in a hierarchical learning framework. We present HiTKG, a hierarchical transformer-based graph walker that leverages multiscale inputs to make precise and flexible predictions on KG paths. Furthermore, we propose a two-hierarchy learning framework that employs two stages to learn both turn-level (short-term) and global-level (long-term) conversation goals. Specifically, at the first stage, HiTKG is trained in a supervised fashion to learn how to plan turn-level goal sequences; at the second stage, HiTKG tries to naturally approach the assigned global goal via reinforcement learning. In addition, we propose MetaPath as the backbone method for KG path representation to exploit the entity and relation information concurrently. We further propose Multi-source Decoding Inputs and Output-level Length Head to improve the decoding controllability. Our experiments show that HiTKG achieves a significant improvement in the performance of turn-level goal learning compared with state-of-the-art baselines. Additionally, both automatic and human evaluation prove the effectiveness of the two-hierarchy learning framework for both short-term and long-term goal planning.

The Linear Threshold Model is a widely used model that describes how information diffuses through a social network. According to this model, an individual adopts an idea or product after the proportion of their neighbors who have adopted it reaches a certain threshold. Typical applications of the Linear Threshold Model assume that thresholds are either the same for all network nodes or randomly distributed, even though some people may be more susceptible to peer pressure than others. To address individual-level differences, we propose causal inference methods for estimating individual thresholds that can more accurately predict whether and when individuals will be affected by their peers. We introduce the concept of heterogeneous peer effects and develop a Structural Causal Model which corresponds to the Linear Threshold Model and supports heterogeneous peer effect identification and estimation. We develop two algorithms for individual threshold estimation, one based on causal trees and one based on causal meta-learners. Our experimental results on synthetic and real- world datasets show that our proposed models can better predict individual-level thresholds in the Linear Threshold Model and thus more precisely predict which nodes will get activated over time.

We initiate the study of the heterogeneous facility location problem with limited resources. We mainly focus on the fundamental case where a set of agents are positioned in the line segment [0,1] and have approval preferences over two available facilities. A mechanism takes as input the positions and the preferences of the agents, and chooses to locate a single facility based on this information. We study mechanisms that aim to maximize the social welfare (the total utility the agents derive from facilities they approve), under the constraint of incentivizing the agents to truthfully report their positions and preferences. We consider three different settings depending on the level of agent-related information that is public or private. For each setting, we design deterministic and randomized strategyproof mechanisms that achieve a good approximation of the optimal social welfare, and complement these with nearly-tight impossibility results.

Twitter bot detection has become an important and challenging task to combat misinformation and protect the integrity of the online discourse. State-of-the-art approaches generally leverage the topological structure of the Twittersphere, while they neglect the heterogeneity of relations and influence among users. In this paper, we propose a novel bot detection framework to alleviate this problem, which leverages the topological structure of user-formed heterogeneous graphs and models varying influence intensity between users. Specifically, we construct a heterogeneous information network with users as nodes and diversified relations as edges. We then propose relational graph transformers to model heterogeneous influence between users and learn node representations. Finally, we use semantic attention networks to aggregate messages across users and relations and conduct heterogeneity-aware Twitter bot detection. Extensive experiments demonstrate that our proposal outperforms state-of-the-art methods on a comprehensive Twitter bot detection benchmark. Additional studies also bear out the effectiveness of our proposed relational graph transformers, semantic attention networks and the graph-based approach in general.

In hedonic games, a set of n agents, having preferences over all possible coalition structures, needs to agree on a stable outcome. In this work, we initiate the study of hedonic games with fixed-size coalitions, where the set of possible coalition structures is restricted as follows: there are k coalitions, each coalition has a fixed size, and the sum of the sizes of all coalitions equals n. We focus on the basic model of additively separable hedonic games with symmetric preferences, where an agent's preference is captured by a utility function which sums up a contribution due to any other agent in the same coalition. In this setting, an outcome is stable if no pair of agents can exchange coalitions and improve their utilities. Conditioned on the definition of improvement, three stability notions arise: swap stability under transferable utilities, which requires to improve the sum of the utilities of both agents, swap stability, which requires to improve the utility of one agent without decreasing the utility of the other one, and strict swap stability, requiring to improve the utilities of both agents simultaneously. We analyse the fundamental questions of existence, complexity and efficiency of stable outcomes, and that of complexity of a social optimum.

Hedonic diversity games are a variant of the classical Hedonic games designed to better model a variety of questions concerning diversity and fairness. Previous works mainly targeted the case with two diversity classes (represented as colors in the model) and provided a set of initial complexity-theoretic and existential results concerning Nash and Individually stable outcomes. Here, we design new algorithms accompanied with lower bounds which provide a full parameterized-complexity picture for computing Nash and Individually stable outcomes with respect to the most natural parameterizations of the problem. Crucially, our results hold for general Hedonic diversity games where the number of colors is not necessarily restricted to two, and show that---apart from two trivial cases---a necessary condition for tractability in this setting is that the number of colors is bounded by the parameter. Moreover, for the special case of two colors we resolve an open question asked in previous work~(Boehmer and Elkind, AAAI 2020).

Multiple medical institutions collaboratively training a model using federated learning (FL) has become a promising solution for maximizing the potential of data-driven models, yet the non-independent and identically distributed (non-iid) data in medical images is still an outstanding challenge in real-world practice. The feature heterogeneity caused by diverse scanners or protocols introduces a drift in the learning process, in both local (client) and global (server) optimizations, which harms the convergence as well as model performance. Many previous works have attempted to address the non-iid issue by tackling the drift locally or globally, but how to jointly solve the two essentially coupled drifts is still unclear. In this work, we concentrate on handling both local and global drifts and introduce a new harmonizing framework called HarmoFL. First, we propose to mitigate the local update drift by normalizing amplitudes of images transformed into the frequency domain to mimic a unified imaging setting, in order to generate a harmonized feature space across local clients. Second, based on harmonized features, we design a client weight perturbation guiding each local model to reach a flat optimum, where a neighborhood area of the local optimal solution has a uniformly low loss. Without any extra communication cost, the perturbation assists the global model to optimize towards a converged optimal solution by aggregating several local flat optima. We have theoretically analyzed the proposed method and empirically conducted extensive experiments on three medical image classification and segmentation tasks, showing that HarmoFL outperforms a set of recent state-of-the-art methods with promising convergence behavior. Code is available at: https://github.com/med-air/HarmoFL

The right to erasure requires removal of a user's information from data held by organizations, with rigorous interpretations extending to downstream products such as learned models. Retraining from scratch with the particular user's data omitted fully removes its influence on the resulting model, but comes with a high computational cost. Machine "unlearning" mitigates the cost incurred by full retraining: instead, models are updated incrementally, possibly only requiring retraining when approximation errors accumulate. Rapid progress has been made towards privacy guarantees on the indistinguishability of unlearned and retrained models, but current formalisms do not place practical bounds on computation. In this paper we demonstrate how an attacker can exploit this oversight, highlighting a novel attack surface introduced by machine unlearning. We consider an attacker aiming to increase the computational cost of data removal. We derive and empirically investigate a poisoning attack on certified machine unlearning where strategically designed training data triggers complete retraining when removed.

Handwritten mathematical expression recognition aims to automatically generate LaTeX sequences from given images. Currently, attention-based encoder-decoder models are widely used in this task. They typically generate target sequences in a left-to-right (L2R) manner, leaving the right-to-left (R2L) contexts unexploited. In this paper, we propose an Attention aggregation based Bi-directional Mutual learning Network (ABM) which consists of one shared encoder and two parallel inverse decoders (L2R and R2L). The two decoders are enhanced via mutual distillation, which involves one-to-one knowledge transfer at each training step, making full use of the complementary information from two inverse directions. Moreover, in order to deal with mathematical symbols in diverse scales, an Attention Aggregation Module (AAM) is proposed to effectively integrate multi-scale coverage attentions. Notably, in the inference phase, given that the model already learns knowledge from two inverse directions, we only use the L2R branch for inference, keeping the original parameter size and inference speed. Extensive experiments demonstrate that our proposed approach achieves the recognition accuracy of 56.85 % on CROHME 2014, 52.92 % on CROHME 2016, and 53.96 % on CROHME 2019 without data augmentation and model ensembling, substantially outperforming the state-of-the-art methods. The source code is available in https://github.com/XH-B/ABM.

This work studies the influence of slice permutations on tensor recovery, which is derived from a reasonable assumption about algorithm, i.e. changing data order should not affect the effectiveness of the algorithm. However, as we will discussed in this paper, this assumption is not satisfied by tensor recovery under some cases. We call this interesting problem as Slice Permutations Variability (SPV) in tensor recovery. In this paper, we discuss SPV of several key tensor recovery problems theoretically and experimentally. The obtained results show that there is a huge gap between results by tensor recovery using tensor with different slices sequences. To overcome SPV in tensor recovery, we develop a novel tensor recovery algorithm by Minimum Hamiltonian Circle for SPV (TRSPV) which exploits a low dimensional subspace structures within data tensor more exactly. To the best of our knowledge, this is the first work to discuss and effectively solve the SPV problem in tensor recovery. The experimental results demonstrate the effectiveness of the proposed algorithm in eliminating SPV in tensor recovery.

Current representation learning methods for whole slide image (WSI) with pyramidal resolutions are inherently homogeneous and flat, which cannot fully exploit the multiscale and heterogeneous diagnostic information of different structures for comprehensive analysis. This paper presents a novel graph neural network-based multiple instance learning framework (i.e., H^2-MIL) to learn hierarchical representation from a heterogeneous graph with different resolutions for WSI analysis. A heterogeneous graph with the “resolution” attribute is constructed to explicitly model the feature and spatial-scaling relationship of multi-resolution patches. We then design a novel resolution-aware attention convolution (RAConv) block to learn compact yet discriminative representation from the graph, which tackles the heterogeneity of node neighbors with different resolutions and yields more reliable message passing. More importantly, to explore the task-related structured information of WSI pyramid, we elaborately design a novel iterative hierarchical pooling (IHPool) module to progressively aggregate the heterogeneous graph based on scaling relationships of different nodes. We evaluated our method on two public WSI datasets from the TCGA project, i.e., esophageal cancer and kidney cancer. Experimental results show that our method clearly outperforms the state-of-the-art methods on both tumor typing and staging tasks.

Recently, deep unfolding networks (DUNs) based on optimization algorithms have received increasing attention, and their high efficiency has been confirmed by many experimental and theoretical results. Since this type of networks combines model-based traditional optimization algorithms, they have high interpretability. In addition, ordinary differential equations (ODEs) are often used to explain deep neural networks, and provide some inspiration for designing innovative network models. In this paper, we transform DUNs into first-order ODE forms, and propose a high-order numerical architecture for ODE-inspired deep unfolding networks. To the best of our knowledge, this is the first work to establish the relationship between DUNs and ODEs. Moreover, we take two representative DUNs as examples, apply our architecture to them and design novel DUNs. In theory, we prove the existence, uniqueness of the solution and convergence of the proposed network, and also prove that our network obtains a fast linear convergence rate. Extensive experiments verify the effectiveness and advantages of our architecture.

The demands of the modern world are increasingly responsible for causing psychological burdens and bringing adverse impacts on our mental health. As a result, neural conversational agents with empathetic responding and distress management capabilities have recently gained popularity. However, existing end-to-end empathetic conversational agents often generate generic and repetitive empathetic statements such as "I am sorry to hear that", which fail to convey specificity to a given situation. Due to the lack of controllability in such models, they also impose the risk of generating toxic responses. Chatbots leveraging reasoning over knowledge graphs is seen as an efficient and fail-safe solution over end-to-end models. However, such resources are limited in the context of emotional distress. To address this, we introduce HEAL, a knowledge graph developed based on 1M distress narratives and their corresponding consoling responses curated from Reddit. It consists of 22K nodes identifying different types of stressors, speaker expectations, responses, and feedback types associated with distress dialogues and forms 104K connections between different types of nodes. Each node is associated with one of 41 affective states. Statistical and visual analysis conducted on HEAL reveals emotional dynamics between speakers and listeners in distress-oriented conversations and identifies useful response patterns leading to emotional relief. Automatic and human evaluation experiments show that HEAL's responses are more diverse, empathetic, and reliable compared to the baselines.

The goal of the crime forecasting problem is to predict different types of crimes for each geographical region (like a neighborhood or censor tract) in the near future. Since nearby regions usually have similar socioeconomic characteristics which indicate similar crime patterns, recent state-of-the-art solutions constructed a distance-based region graph and utilized Graph Neural Network (GNN) techniques for crime forecasting, because the GNN techniques could effectively exploit the latent relationships between neighboring region nodes in the graph if the edges reveal high dependency or correlation. However, this distance-based pre-defined graph can not fully capture crime correlation between regions that are far from each other but share similar crime patterns. Hence, to make a more accurate crime prediction, the main challenge is to learn a better graph that reveals the dependencies between regions in crime occurrences and meanwhile captures the temporal patterns from historical crime records. To address these challenges, we propose an end-to-end graph convolutional recurrent network called HAGEN with several novel designs for crime prediction. Specifically, our framework could jointly capture the crime correlation between regions and the temporal crime dynamics by combining an adaptive region graph learning module with the Diffusion Convolution Gated Recurrent Unit (DCGRU). Based on the homophily assumption of GNN (i.e., graph convolution works better where neighboring nodes share the same label), we propose a homophily-aware constraint to regularize the optimization of the region graph so that neighboring region nodes on the learned graph share similar crime patterns, thus fitting the mechanism of diffusion convolution. Empirical experiments and comprehensive analysis on two real-world datasets showcase the effectiveness of HAGEN.

Semi-supervised learning is a challenging problem which aims to construct a model by learning from limited labeled examples. Numerous methods for this task focus on utilizing the predictions of unlabeled instances consistency alone to regularize networks. However, treating labeled and unlabeled data separately often leads to the discarding of mass prior knowledge learned from the labeled examples. In this paper, we propose a novel method for semi-supervised semantic segmentation named GuidedMix-Net, by leveraging labeled information to guide the learning of unlabeled instances. Specifically, GuidedMix-Net employs three operations: 1) interpolation of similar labeled-unlabeled image pairs; 2) transfer of mutual information; 3) generalization of pseudo masks. It enables segmentation models can learning the higher-quality pseudo masks of unlabeled data by transfer the knowledge from labeled samples to unlabeled data. Along with supervised learning for labeled data, the prediction of unlabeled data is jointly learned with the generated pseudo masks from the mixed data. Extensive experiments on PASCAL VOC 2012, and Cityscapes demonstrate the effectiveness of our GuidedMix-Net, which achieves competitive segmentation accuracy and significantly improves the mIoU over 7$%$ compared to previous approaches.

Local image feature matching under large appearance, viewpoint, and distance changes is challenging yet important. Conventional methods detect and match tentative local features across the whole images, with heuristic consistency checks to guarantee reliable matches. In this paper, we introduce a novel Overlap Estimation method conditioned on image pairs with TRansformer, named OETR, to constrain local feature matching in the commonly visible region. OETR performs overlap estimation in a two step process of feature correlation and then overlap regression. As a preprocessing module, OETR can be plugged into any existing local feature detection and matching pipeline, to mitigate potential view angle or scale variance. Intensive experiments show that OETR can boost state of the art local feature matching performance substantially, especially for image pairs with small shared regions. The code will be publicly available at https://github.com/AbyssGaze/OETR.

The fairness in machine learning is getting increasing attention, as its applications in different fields continue to expand and diversify. To mitigate the discriminated model behaviors between different demographic groups, we introduce a novel post-processing method to optimize over multiple fairness constraints through group-aware threshold adaptation. We propose to learn adaptive classification thresholds for each demographic group by optimizing the confusion matrix estimated from the probability distribution of a classification model output. As we only need an estimated probability distribution of model output instead of the classification model structure, our post-processing model can be applied to a wide range of classification models and improve fairness in a model-agnostic manner and ensure privacy. This even allows us to post-process existing fairness methods to further improve the trade-off between accuracy and fairness. Moreover, our model has low computational cost. We provide rigorous theoretical analysis on the convergence of our optimization algorithm and the trade-off between accuracy and fairness. Our method theoretically enables a better upper bound in near optimality than previous method under the same condition. Experimental results demonstrate that our method outperforms state-of-the-art methods and obtains the result that is closest to the theoretical accuracy-fairness trade-off boundary.

Effectively integrating knowledge into end-to-end task-oriented dialog systems remains a challenge. It typically requires incorporation of an external knowledge base (KB) and capture of the intrinsic semantics of the dialog history. Recent research shows promising results by using Sequence-to-Sequence models, Memory Networks, and even Graph Convolutional Networks. However, current state-of-the-art models are less effective at integrating dialog history and KB into task-oriented dialog systems in the following ways: 1. The KB representation is not fully context-aware. The dynamic interaction between the dialog history and KB is seldom explored. 2. Both the sequential and structural information in the dialog history can contribute to capturing the dialog semantics, but they are not studied concurrently. In this paper, we propose a novel Graph Memory Network (GMN) based Seq2Seq model, GraphMemDialog, to effectively learn the inherent structural information hidden in dialog history, and to model the dynamic interaction between dialog history and KBs. We adopt a modified graph attention network to learn the rich structural representation of the dialog history, whereas the context-aware representation of KB entities are learnt by our novel GMN. To fully exploit this dynamic interaction, we design a learnable memory controller coupled with external KB entity memories to recurrently incorporate dialog history context into KB entities through a multi-hop reasoning mechanism. Experiments on three public datasets show that our GraphMemDialog model achieves state-of-the-art performance and outperforms strong baselines by a large margin, especially on datatests with more complicated KB information.