AAAI 2025 论文列表

Few-shot Named Entity Recognition (NER) spotlights the tag of novel entity types in data-limited scenarios or lower-resource settings. Advances with Pre-trained Language Models (PLMs), including BERT, GPT, and their variants, have driven tremendous strategies to leverage context-dependent representations and exploit predefined relational cues, yielding significant gains in witnessing unseen entities. Nevertheless, a fundamental issue exists in prior efforts regarding their susceptibility to adversarial attacks in the intricate semantic environment. This vulnerability undermines the robustness of semantic representations, exacerbating the challenge of accurate entity identification, especially when transitioning across domains. To this end, we propose an Adversity-aware Augment Learning (AAL) solution for the few-shot NER task, dedicated to retrieving and reinforcing entity prototypes resilient to adversarial inference, thereby enhancing cross-domain semantic coherence. In particular, AAL employs a two-stage paradigm consisting of training and fine-tuning. The process initiates with augmentation learning by leveraging two kinds of prompt learning schemes, then identifies prototypes under the guidance of a variational manner. Furthermore, we devise a domain-oriented prototype refinement to optimize prototype learning under conditions of uncertainty attack, facilitating the effective transfer of common knowledge from source to target domains. The experimental results, encompassing the few-shot NER datasets under both certainty and uncertainty conditions, affirm the superiority of the proposed AAL over several representative baselines, particularly its capability against adversarial attacks.

Backdoor attacks and adversarial attacks are two major security threats to deep neural networks (DNNs), with the former one is a training-time data poisoning attack that aims to implant backdoor triggers into models by injecting trigger patterns into training samples, and the latter one is a testing-time attack trying to generate adversarial examples (AEs) from benign images to mislead a well-trained model. While previous works generally treat these two attacks separately, the inherent connection between these two attacks is rarely explored. In this paper, we focus on bridging backdoor and adversarial attacks and observe two intriguing phenomena when applying adversarial attacks on an infected model implanted with backdoors: 1) the sample is harder to be turned into an AE when the trigger is presented; 2) the AEs generated from backdoor samples are highly likely to be predicted as its true labels. Inspired by these observations, we proposed a novel backdoor defense method, dubbed Adversarial-Inspired Backdoor Defense (AIBD), to isolate the backdoor samples by leveraging a progressive top-q scheme and break the correlation between backdoor samples and their target labels using adversarial labels. Through extensive experiments on various datasets against six state-of-the-art backdoor attacks, the AIBD-trained models on poisoned data demonstrate superior performance over the existing defense methods.

Acute lymphoblastic leukemia is a childhood cancer prevalent worldwide, which can prove fatal within weeks or months. However, current diagnosis models based on machine learning and deep learning methods fail to consider device noise (pixel-level perturbations) and rotation/translation (spatial-transformed perturbations), which can undermine the model's robustness. Adversarial training is a potential solution to this issue. This paper presents a hybrid perturbation adversarial training (HPAT) strategy that leverages two types of adversarial samples: pixel-level adversarial samples and spatial adversarial samples. This work generates these hybrid adversarial samples through Projected Gradient Descent (PGD) in couple with spatial transformation based on the Bayesian optimization (STBO) algorithm, respectively. This work introduced the Mixed Batch Normalization (MixBN) module to handle both adversarial samples and clean samples, alleviating the problem of clean accuracy degradation due to adversarial training. The proposed hybrid adversarial training strategy is tested on the public acute lymphoblastic leukemia dataset and found that it outperformed existing acute lymphoblastic cell classification models.

Graph Neural Networks (GNNs) have been shown vulnerable to graph adversarial attacks. Current robust graph representation learning methods mainly defend against graph structure attack, and improves performance of GNNs. However node feature in graph can been easily attacked in reality. The joint defense on graph structure and feature dual attacks remains challenging yet less studied. To fulfill this gap, we propose Adversarial Contrastive Graph Masked AutoEncoder (ACGMAE) to defend against graph structure and feature dual attacks. ACGMAE employs adversarial feature masking for reconstructing node feature to mitigate the influence of feature attack. ACGMAE employs contrastive learning on kNN graph and attacked graph, considers neighbor nodes as positive samples, and further calculates their probabilities being true positive to mitigate the effect of adversarial edges. Extensive experiments on node classification and clustering demonstrate the effectiveness of the proposed ACGMAE especially under graph structure and feature dual attacks.

With the advancement of graph representation learning, self-supervised graph contrastive learning (GCL) has emerged as a key technique in the field. In GCL, positive and negative samples are generated through data augmentation. While recent works have introduced model-based methods to enhance positive graph augmentations, they often overlook the importance of negative samples, relying instead on rule-based methods that can fail to capture meaningful graph patterns. To address this issue, we propose a novel model-based adversarial contrastive graph augmentation (ACGA) method that automatically generates both positive graph samples with minimal sufficient information and hard negative graph samples. Additionally, we provide a theoretical framework to analyze the process of positive and negative graph augmentation in self-supervised GCL. We evaluate our ACGA method through extensive experiments on representative benchmark datasets, and the results demonstrate that ACGA outperforms state-of-the-art baselines.

Event cameras, known for their low latency and high dynamic range, show great potential in pedestrian detection applications. However, while recent research has primarily focused on improving detection accuracy, the robustness of event-based visual models against physical adversarial attacks has received limited attention. For example, adversarial physical objects, such as specific clothing patterns or accessories, can exploit inherent vulnerabilities in these systems, leading to misdetections or misclassifications. This study is the first to explore physical adversarial attacks on event-driven pedestrian detectors, specifically investigating whether certain clothing patterns worn by pedestrians can cause these detectors to fail, effectively rendering them unable to detect the person. To address this, we developed an end-to-end adversarial framework in the digital domain, framing the design of adversarial clothing textures as a 2D texture optimization problem. By crafting an effective adversarial loss function, the framework iteratively generates optimal textures through backpropagation. Our results demonstrate that the textures identified in the digital domain possess strong adversarial properties. Furthermore, we translated these digitally optimized textures into physical clothing and tested them in real-world scenarios, successfully demonstrating that the designed textures significantly degrade the performance of event-based pedestrian detection models. This work highlights the vulnerability of such models to physical adversarial attacks.

Recent advancements in neuroscience research have propelled the development of Spiking Neural Networks (SNNs), which not only have the potential to further advance neuroscience research but also serve as an energy-efficient alternative to Artificial Neural Networks (ANNs) due to their spike-driven characteristics. However, previous studies often overlooked the multiscale information and its spatiotemporal correlation between event data, leading SNN models to approximate each frame of input events as static images. We hypothesize that this oversimplification significantly contributes to the performance gap between SNNs and traditional ANNs. To address this issue, we have designed a Spiking Multiscale Attention (SMA) module that captures multiscale spatiotemporal interaction information. Furthermore, we developed a regularization method named Attention ZoneOut (AZO), which utilizes spatiotemporal attention weights to reduce the model's generalization error through pseudo-ensemble training. Our approach has achieved state-of-the-art results on mainstream neuromorphic datasets. Additionally, we have reached a performance of 77.1\% on the Imagenet-1K dataset using a 104-layer ResNet architecture enhanced with SMA and AZO. This achievement confirms the state-of-the-art performance of SNNs with non-transformer architectures and underscores the effectiveness of our method in bridging the performance gap between SNN models and traditional ANN models.

Diffusion-based molecular graph generative models have achieved significant success in template-free, single-step retrosynthesis prediction. However, these models typically generate reactants from scratch, often overlooking the fact that the scaffold of a product molecule typically remains unchanged during chemical reactions. To leverage this useful observation, we introduce a retrieval-augmented molecular graph generation framework. Our framework comprises three key components: a retrieval component that identifies similar molecules for the given product, an integration component that learns valuable clues from these molecules about which part of the product should remain unchanged, and a base generative model that is prompted by these clues to generate the corresponding reactants. We explore various design choices for critical and under-explored aspects of this framework and instantiate it as the Retrieval-Augmented RetroBridge (RARB). RARB demonstrates state-of-the-art performance on standard benchmarks, achieving a 14.8% relative improvement in top-1 accuracy over its base generative model, highlighting the effectiveness of retrieval augmentation. Additionally, RARB excels in handling out-of-distribution molecules, and its advantages remain significant even with smaller models or fewer denoising steps. These strengths make RARB highly valuable for real-world retrosynthesis applications, where extrapolation to novel molecules and high-throughput prediction are essential.

Part of a university initiative supporting responsible AI for social empowerment and education, the project-based RAICA (Responsible AI for Computational Action) curriculum supports middle/high school learners and novice AI literacy teachers use AI creatively for good. This paper offers a rare example of design-based implementation research (DBIR) in AI education across widely varied contexts, provides fine grain implementation data that contributes to a foundation for evaluating effectiveness and expanding access. We present a novel approach to analyzing fidelity of implementation data from RAICA’s computer vision module beta-test. Twelve educators working with ~282 students across nine pilot sites in four countries used a bespoke fidelity of implementation data collection tool (pre-made comment prompts in a Google Docs version of the teacher guide) to provide 236 qualitative responses about AI literacy and responsible design activities, plus 111 ordinal ratings of embedded teacher supports. Analyses revealed that while the curriculum was generally implemented as designed, educators frequently made modifications. Although most changes produced practical insights for improved curriculum design, others helped the design team anticipate and prevent changes that could obscure learning objectives and hinder outcomes. We discuss the pedagogical, design, and research implications of these findings for effective AI teaching/learning in diverse settings.

Synthesizing electronic health records (EHR) is essential for addressing data scarcity, bias, and fairness in healthcare models. EHR data are inherently multimodal and sequential, encompassing structured codes, clinical notes, medical images, and irregular time intervals. Traditional generative models like GANs and VAEs struggle to capture these complexities, while diffusion-based models offer improvements but remain limited to task-specific applications. To address these challenges, two diffusion-based models, MedDiffusion and EHRPD, have been developed. MedDiffusion enhances health risk prediction by generating synthetic patient data and capturing visit-level relationships, while EHRPD generates sequential, multimodal EHR data, incorporating temporal interval estimation to improve diversity and fidelity. Future work aims to overcome limitations in multimodal data generation by developing a generalized model capable of handling diverse modalities simultaneously, expanding the applicability of EHR data generation across healthcare tasks.

Loss functions play a pivotal role in optimizing recommendation models. Among various loss functions, Softmax Loss (SL) and Cosine Contrastive Loss (CCL) are particularly effective. Their theoretical connections and differences warrant in-depth exploration. This work conducts comprehensive analyses of these losses, yielding significant insights: 1) Common strengths --- both can be viewed as augmentations of traditional losses with Distributional Robust Optimization (DRO), enhancing robustness to distributional shifts; 2) Respective limitations --- stemming from their use of different distribution distance metrics in DRO optimization, SL exhibits high sensitivity to false negative instances, whereas CCL suffers from low data utilization. To address these limitations, this work proposes a new loss function, DrRL, which generalizes SL and CCL by leveraging Rényi-divergence in DRO optimization. DrRL incorporates the advantageous structures of both SL and CCL, and can be demonstrated to effectively mitigate their limitations. Extensive experiments have been conducted to validate the superiority of DrRL on both recommendation accuracy and robustness.

Integrating Large Language Models (LLMs) into software engineering unlocks new opportunities to automate manual processes but raises challenges around reliability, safety, and scalability. My research centers on this synergy, with two key objectives: first, harnessing LLMs to solve software engineering tasks traditionally dependent on labor-intensive, domain-specific methods, and second, applying robust software engineering principles to improve LLM safety and performance. This dual focus creates a powerful feedback loop, where LLMs drive innovation while engineering rigor ensures these systems meet the high standards required for real-world applications.

The exponential growth of unstructured medical data presents a unique opportunity and challenge for advancing healthcare. Traditional methods struggle to extract meaningful insights from this complex data due to its inherent noise, ambiguity, and heterogeneity. To address these limitations, we propose a novel hybrid approach that integrates Knowledge Graphs (KGs) with Large Language Models (LLMs) within a Retrieval-Augmented Generation (RAG) framework. By leveraging the structured knowledge of KGs and the contextual understanding of LLMs, we aim to improve the precision of feature extraction and disease progression modeling. Our research focuses on refining the KG representation through advanced entity extraction and relation extraction techniques, ensuring that the KG accurately captures the semantic nuances and temporal dynamics of medical data. By integrating this enhanced KG with the RAG framework, we can derive more precise and informative insights for clinical decision-making.

Image Aesthetic Assessment (IAA) is a vital and intricate task that entails analyzing and assessing an image's aesthetic values, and identifying its highlights and areas for improvement. Traditional methods of IAA often concentrate on a single aesthetic task and suffer from inadequate labeled datasets, thus impairing in-depth aesthetic comprehension. Despite efforts to overcome this challenge through the application of Multi-modal Large Language Models (MLLMs), such models remain underdeveloped for IAA purposes. To address this, we propose a comprehensive aesthetic MLLM capable of nuanced aesthetic insight. Central to our approach is an innovative multi-scale text-guided self-supervised learning technique. This technique features a multi-scale feature alignment module and capitalizes on a wealth of unlabeled data in a self-supervised manner to structurally and functionally enhance aesthetic ability. The empirical evidence indicates that accompanied with extensive instruct-tuning, our model sets new state-of-the-art benchmarks across multiple tasks, including aesthetic scoring, aesthetic commenting, and personalized image aesthetic assessment. Remarkably, it also demonstrates zero-shot learning capabilities in the emerging task of aesthetic suggesting. Furthermore, for personalized image aesthetic assessment, we harness the potential of in-context learning and showcase its inherent advantages.

Human feedback in generative systems is a highly active frontier of research that aims to improve the quality of generated content and align it with subjective preferences. Existing efforts predominantly focus on text-only large language models (LLMs) or text-based image generation, while cross-modal generation between audio and text remains largely unexplored. Moreover, there is currently no open-source preference dataset to support the deployment of alignment algorithms in this domain. In this work, we take audio speech translation (AST) and audio captioning (AAC) tasks as examples to explore how to enhance the performance of mainstream audio-based text generation models with limited human annotation. Specifically, we propose an novel framework named IPO that includes a model adversarial sampling concept--human annotators act as referees to determine model outcomes, using these results as pseudo-labels for the corresponding beam search hypotheses. Given these imbalance win-loss results, IPO effectively enable the two models to update interactively to win the next round of adversarial sampling. We conduct both subjective and objective evaluations to demonstrate the alignment benefits of IPO and its enhancement on model perception and generation capacities. On both AAC and AST, a few hundreds of annotations significantly enhance the weak model, and the strong model can also be encouraged to achieve new state-of-the-art results in terms of different objective metrics. Additionally, we show the extensibility of IPO by applying it to the reverse task of text-to-speech generation, improving the robustness of system on unseen reference speaker.

Artificial intelligence has made remarkable progress in reasoning over complex, structured, multimodal, and multilingual data, addressing critical challenges in domains such as finance and healthcare. This abstract underscores key advancements in tabular reasoning, temporal analysis, and structured multimodal reasoning. Key contributions include the development of TempTabQA, a benchmark for temporal question answering, along with novel methods for enhancing temporal reasoning in large language models (LLMs). Additionally, a framework for evaluating mathematical reasoning in financial documents has been introduced, establishing robust techniques for interpreting time-sensitive and quantitative data. Building on these foundations, we have developed hybrid SQL-text adaptive reasoning models (H-STAR) and knowledge-aware reasoning techniques for semi-structured tables (MMTabQA), enabling precise and efficient handling of complex queries. In the vision-language domain, our contributions include advancements in spatial reasoning for geographic data (MAPWise), methods to improve robustness in chart interpretation (FlowVQA), and evaluations of LLMs’ ability to understand visual data, such as charts. Furthermore, we have addressed challenges in multilingual and cross-modal robustness through innovations such as multilingual table synchronization (InfoSync), concurrent robustness evaluations across languages and modalities, and numerical reasoning in tabular data. Our work aims to enhance reasoning on dynamically evolving data using hybrid LLM-SQL queries, symbolic query generation, and multi-table retrieval techniques. We also plan to tackle challenges in interpreting hierarchical table structures, analyzing multiple complex chart types, and exploring diverse map types, while advancing real-world multimodal data analysis. Additionally, we plan to improve table generation in both closed/open-book scenarios and refine evaluation frameworks for structured tasks. These advancements demonstrate the potential of AI in tackling complex, multimodal data and delivering impactful real-world solutions.

When the current physical adversarial patches cannot deceive thermal infrared detectors, the existing techniques implement adversarial attacks from scratch, such as digital patch generation, material production, and physical deployment. Besides, it is difficult to finely regulate infrared radiation. To address these issues, this paper designs an adversarial thermal display (AdvDisplay ) by assembling thermoelectric coolers (TECs) as an array. Specifically, to reduce the gap between patches in the physical and digital worlds and decrease the power of AdvDisplay device, heat transfer loss and electric power loss are designed to guide the patch optimization. In addition, a precise temperature control scheme for AdvDisplay is proposed based on proportional-integral-derivative (PID) control. Due to the accurate temperature regulation and the reusability of AdvDisplay , our method is able to improve the attack success rate and the efficiency of physical deployments. Extensive experimental results indicate that the proposed method possesses superior adversarial effectiveness compared to other methods and demonstrates strong robustness in physical attacks.

Multi-label Learning with Partial Labels (ML-PL) learns from training data, where each sample is annotated with part of positive labels while leaving the rest of positive labels unannotated. Existing methods mainly focus on extending multi-label losses to estimate unannotated labels, further inducing a missing-robust network. However, training with single network could lead to confirmation bias (i.e., the model tends to confirm its mistakes). To tackle this issue, we propose a novel learning paradigm termed Co-Label Selection (CLS), where two networks feed forward all data and cooperate in a co-training manner for critical label selection. Different from traditional co-training based methods that networks select confident samples for each other, we start from a new perspective that two networks are encouraged to remove false-negative labels while keep training samples reserved. Meanwhile, considering the extreme positive-negative label imbalance in ML-PL that leads the model to focus on negative labels, we enforce the model to concentrate on positive labels by abandoning non-informative negative labels to alleviate such issue. By shifting the cooperation strategy from "Sample Selection'' to "Label Selection'', CLS avoids directly dropping samples and reserves training data in most extent, thus enhancing the utilization of supervised signals and the generalization of the learning model. Empirical results performed on various multi-label datasets demonstrate that our CLS is significantly superior to other state-of-the-art methods.

Predicting Click-Through Rates is a crucial function within recommendation and advertising platforms, as the output of CTR prediction determines the order of items shown to users. The Embedding and MLP paradigm has become a standard approach for industrial recommendation systems and has been widely deployed. However, this paradigm suffers from cold-start problems, where there is either no or only limited user action data available, leading to poorly learned ID embeddings. The cold-start problem hampers the performance of new items. To address this problem, we design a novel diffusion model to generate a warmed-up embedding for new items. Specifically, we define a novel diffusion process between the ID embedding space and the side information space. In addition, we can derive a sub-sequence from the diffusion steps to expedite training, given that our diffusion model is non-Markovian. Our diffusion model is supervised by both the variational inference and binary cross-entropy objectives, enabling it to generate warmed-up embeddings for items in both the cold-start and warm-up phases. Additionally, we have conducted extensive experiments on three recommendation datasets. The results confirmed the effectiveness of our approach.