AAAI 2025 Papers

The storage and recall of factual associations in auto-regressive transformer language models (LMs) have drawn a great deal of attention, inspiring knowledge editing by directly modifying the located model weights. Most editing works achieve knowledge editing under the guidance of existing interpretations of knowledge recall that mainly focus on subject knowledge. However, these interpretations are seriously flawed, neglecting relation information and leading to the *over-generalizing* problem for editing. In this work, we discover a novel relation-focused perspective to interpret the knowledge recall of transformer LMs during inference and apply it on single knowledge editing to avoid over-generalizing. Experimental results on the dataset supplemented with a new R-Specificity criterion demonstrate that our editing approach significantly alleviates over-generalizing while remaining competitive on other criteria, breaking the domination of subject-focused editing for future research.

We present a Reinforcement Learning Platform for Adversarial Black-box untargeted and targeted attacks, RLAB, that allows users to select from various distortion filters to create adversarial examples. The platform uses a Reinforcement Learning agent to add minimum distortion to input images while still causing misclassification by the target model. The agent uses a novel dual-action method to explore the input image at each step to identify sensitive regions for adding distortions while removing noises that have less impact on the target model. This dual action leads to faster and more efficient convergence of the attack. The platform can also be used to measure the robustness of image classification models against specific distortion types. Also, retraining the model with adversarial samples significantly improved robustness when evaluated on benchmark datasets. The proposed platform outperforms state-of-the-art methods in terms of the average number of queries required to cause misclassification. This advances trustworthiness with a positive social impact.

Carefully selecting clients to participate in aggregation can assist the global model in achieving better performance. However, existing research on federated heterogeneous graph learning (FHGL) has shown limited attention to the client selection (CS) problem. Current CS algorithms face challenges in accurately evaluating client contributions and selecting appropriate participants in the context of FHGL, leading to a dilemma between convergence and accuracy. In this paper, we propose a Reinforcement Active client selection based Federated Heterogeneous Graph Learning (RAFHGL), which precisely evaluates the importance of local heterogeneous graph data and selects high-contributing clients for aggregation. RAFHGL employs an active learning agent to select representative nodes for local training. The statistical features of the active scores are used to assess client contributions. A client selection agent then chooses clients conducive to global model convergence for aggregation. To address heterogeneity introduced by sample and client selection, the training process stabilizes by correcting local losses based on data prototypes. Experimental results on 4 publicly available heterogeneous graph datasets show that RAFHGL outperforms existing Client Selection algorithms in federated heterogeneous graph learning scenarios in terms of performance and convergence.

Image forgery detection and localization (IFDL) is of vital importance as forged images can spread misinformation that poses potential threats to our daily life. However, previous methods still struggled to effectively handle forged images processed with diverse forgery operations in real-world scenarios. In this paper, we propose a novel Reinforced Multi-teacher Knowledge Distillation (Re-MTKD) framework for the IFDL task, structured around an encoder-decoder ConvNeXt-UperNet along with Edge-Aware Module, named Cue-Net. First, three Cue-Net models are separately trained for the three main types of image forgeries, i.e., copy-move, splicing and inpainting, which then serve as the multi-teacher models to train the target student model with Cue-Net through self-knowledge distillation. A Reinforced Dynamic Teacher Selection (Re-DTS) strategy is developed to dynamically assign weights to the involved teacher models, which facilitates specific knowledge transfer and enables the student model to effectively learn both the common and specific natures of diverse tampering traces. Extensive experiments demonstrate that, compared with other state-of-the-art methods, the proposed method achieves superior performance on several recently emerged datasets comprised of various kinds of image forgeries.

Representation learning is a powerful tool that enables learning over large multitudes of agents or domains by enforcing that all agents operate on a shared set of learned features. However, many robotics or controls applications that would benefit from collaboration operate in settings with changing environments and goals, whereas most guarantees for representation learning are stated for static settings. Toward rigorously establishing the benefit of representation learning in dynamic settings, we analyze the regret of multi-task representation learning for linear-quadratic control. This setting introduces unique challenges. Firstly, we must account for and balance the misspecification introduced by an approximate representation. Secondly, we cannot rely on the parameter update schemes of single-task online LQR, for which least-squares often suffices, and must devise a novel scheme to ensure sufficient improvement. We demonstrate that for settings where exploration is "benign", the regret of any agent after T timesteps scales with the square root of T/H, where H is the number of agents. In settings with "difficult" exploration, the regret scales as the square root of the input dimension times the parameter dimension multiplied by T, plus a term which scales with T to the three quarters divided by H to the one fifth. In both cases, by comparing to the minimax single-task regret, we see a benefit of a large number of agents. Notably, in the difficult exploration case, by sharing a representation across tasks, the effective task-specific parameter count can often be small. Lastly, we validate the trends we predict.

Real-world optimization problems often involve complex objective functions with costly evaluations. While Bayesian optimization (BO) with Gaussian processes is effective for these challenges, it suffers in high-dimensional spaces due to performance degradation from limited function evaluations. To overcome this, simplification techniques like dimensionality reduction have been employed, yet they often rely on assumptions about the problem characteristics, potentially underperforming when these assumptions do not hold. Trust-region-based methods, which avoid such assumptions, focus on local search but risk stagnation in local optima. In this study, we propose a novel acquisition function, regional expected improvement (REI), designed to enhance trust-region-based BO in medium to high-dimensional settings. REI identifies regions likely to contain the global optimum, improving performance without relying on specific problem characteristics. We provide a theoretical proof that REI effectively identifies optimal trust regions and empirically demonstrate that incorporating REI into trust-region-based BO outperforms conventional BO and other high-dimensional BO methods in medium to high-dimensional real-world problems.

Change captioning aims to describe the differences between two similar images using natural language, significantly aiding in understanding and monitoring changes. This challenging task requires a fine-grained understanding of subtle changes while resisting disturbances like viewpoint shifts and illumination variations. Existing methods often rely solely on global difference features and lack comprehensive alignment of linguistic and visual information, leading to overlooking fine-grained details and generating semantic hallucinated sentences. To address these limitations, we propose the region-aware difference distilling (RDD) network with attribute-guided contrastive regularization (ACR). The RDD uses global difference features to progressively distill regional difference features using learnable vectors, allowing for more precise identification of changed regions. The ACR enhances comprehensive alignment between linguistic and visual information by formulating Nouns-to-Objects (N2O) and Verbs-to-Actions (V2A) alignment losses to regularize the regional difference features. Promising results on three datasets demonstrate that our method outperforms the state-of-the-art change captioning methods.

Rapid advancements in speech synthesis and voice conversion bring convenience but also new security risks, creating an urgent need for effective audio deepfake detection. Although current models perform well, their effectiveness diminishes when confronted with the diverse and evolving nature of real-world deepfakes. To address this issue, we propose a continual learning method named Region-Based Optimization (RegO) for audio deepfake detection. Specifically, we use the Fisher information matrix to measure important neuron regions for real and fake audio detection, dividing them into four regions. First, we directly fine-tune the less important regions to quickly adapt to new tasks. Next, we apply gradient optimization in parallel for regions important only to real audio detection, and in orthogonal directions for regions important only to fake audio detection. For regions that are important to both, we use sample proportion-based adaptive gradient optimization. This region-adaptive optimization ensures an appropriate trade-off between memory stability and learning plasticity. Additionally, to address the increase of redundant neurons from old tasks, we further introduce the Ebbinghaus forgetting mechanism to release them, thereby promoting the model’s ability to learn more generalized discriminative features. Experimental results show our method achieves a 21.3 percent improvement in EER over the state-of-the-art continual learning approach RWM for audio deepfake detection. Moreover, the effectiveness of RegO extends beyond the audio deepfake detection domain, showing potential significance in other tasks, such as image recognition.

Consistency regularization and pseudo-labeling have significantly advanced semi-supervised learning (SSL). Prior works have effectively employed Mixup for consistency regularization in SSL. However, our findings indicate that applying Mixup for consistency regularization may degrade SSL performance by compromising the purity of artificial labels. Moreover, most pseudo-labeling based methods utilize thresholding strategy to exclude low-confidence data, aiming to mitigate confirmation bias; however, this approach limits the utility of unlabeled samples. To address these challenges, we propose RegMixMatch, a novel framework that optimizes the use of Mixup with both high- and low-confidence samples in SSL. First, we introduce semi-supervised RegMixup, which effectively addresses reduced artificial labels purity by using both mixed samples and clean samples for training. Second, we develop a class-aware Mixup technique that integrates information from the top-2 predicted classes into low-confidence samples and their artificial labels, reducing the confirmation bias associated with these samples and enhancing their effective utilization. Experimental results demonstrate that RegMixMatch achieves state-of-the-art performance across various SSL benchmarks.

Graph Neural Networks (GNNs) have exhibited remarkable capabilities for dealing with graph-structured data. However, recent studies have revealed their fragility to adversarial attacks, where imperceptible perturbations to the graph structure can easily mislead predictions. To enhance adversarial robustness, some methods attempt to learn robust representation through improving GNN architectures. Subsequently, another approach suggests that these GNNs might taint feature information and have poor classifier performance, leading to the introduction of Graph Contrastive Learning (GCL) methods to build a refining-classifying pipeline. However, existing methods focus on global-local contrastive strategies, which fails to address the robustness issues inherent in the contexts of adversarial robustness. To address these challenges, we propose a novel paradigm named GRANCE to enhance the robustness of learned representations by shifting the focus to local neighborhoods. Specifically, a dual neighborhood contrastive learning strategy is designed to extract local topological and semantic information. Paired with a neighbor estimator, the strategy can learn robust representations that are resilient to adversarial edges. Additionally, we also provide an improved GNN as classifier. Theoretical analyses provide a stricter lower bound of mutual information, ensuring the convergence of GRANCE. Extensive experiments validate the effectiveness of GRANCE compared to state-of-the-art baselines against various adversarial attacks.

Extracting fine-grained OCR text from aged documents in diacritic languages remains challenging due to unexpected artifacts, time-induced degradation, and lack of datasets. While standalone spell correction approaches have been proposed, they show limited performance for historical documents due to numerous possible OCR error combinations and differences between modern and classical corpus distributions. We propose a method utilizing available content-focused ebooks as a reference base to correct imperfect OCR-generated text, supported by large language models. This technique generates high-precision pseudo-page-to-page labels for diacritic languages, where small strokes pose significant challenges in historical conditions. The pipeline eliminates various types of noise from aged documents and addresses issues such as missing characters, words, and disordered sequences. Our post-processing method, which generated a large OCR dataset of classical Vietnamese books, achieved a mean grading score of 8.72 on a 10-point scale. This outperformed the state-of-the-art transformer-based Vietnamese spell correction model, which scored 7.03, when evaluated on a sampled subset of the dataset. We also trained a baseline OCR model to assess and compare it with well-known engines. Experimental results demonstrate the strength of our baseline model compared to widely used open-source solutions. The resulting dataset will be released publicly to support future studies.

Despite the rapid and substantial advancements in object detection, it continues to face limitations imposed by pre-defined category sets. Current methods for visual grounding primarily focus on how to better leverage the visual backbone to generate text-tailored visual features, which may require adjusting the parameters of the entire model. Besides, some early methods, \ie, matching-based method, build upon and extend the functionality of existing object detectors by enabling them to localize an object based on free-form linguistic expressions, which have good application potential. However, the untapped potential of the matching-based approach has not been fully realized due to inadequate exploration. In this paper, we first analyze the limitations that exist in the current matching-based method (\ie, mismatch problem and complicated fusion mechanisms), and then present a simple yet effective matching-based method, namely RefDetector. To tackle the above issues, we devise a simple heuristic rule to generate proposals with improved referent recall. Additionally, we introduce a straightforward vision-language interaction module that eliminates the need for intricate manually-designed mechanisms. Moreover, we have explored the visual grounding based on the modern detector DETR, and achieved significant performance improvement. Extensive experiments on three REC benchmark datasets, \ie, RefCOCO, RefCOCO+, and RefCOCOg validate the effectiveness of the proposed method.

Two prominent objectives in social choice are utilitarian - maximizing the sum of agents' utilities, and leximin - maximizing the smallest agent's utility, then the second-smallest, etc. Utilitarianism is typically computationally easier to attain but is generally viewed as less fair. This paper presents a general reduction scheme that, given a utilitarian solver, produces a distribution over states (deterministic outcomes) that is leximin in expectation. Importantly, the scheme is robust in the sense that, given an approximate utilitarian solver, it produces a lottery that is approximately-leximin (in expectation) - with the same approximation factor. We apply our scheme to several social choice problems: stochastic allocations of indivisible goods, giveaway lotteries, and fair lotteries for participatory budgeting.

The widespread adoption of Batch Normalization (BN) in contemporary deep neural architectures has demonstrated significant efficacy, particularly in the domain of Unsupervised Domain Adaptation (UDA) for cross-domain applications. Notwithstanding its success, extant BN variants often conflate source and target domain information within identical channels, potentially compromising transferability due to inter-domain feature misalignment. To address this limitation, we introduce Refined Batch Normalization (RBN), a novel normalization paradigm that leverages estimated shift to quantify discrepancies between estimated population statistics and their expected values. Our pivotal observation reveals that estimated shift can accumulate through BN stacking within the network, potentially degrading target domain performance. We elucidate how RBN mitigates this accumulation, thereby enhancing overall system efficacy. The practical implementation of this technique is realized through the RBNBlock, which supplants conventional BN with RBN in the bottleneck architecture of residual networks. Extensive empirical evaluation across diverse cross-domain benchmarks corroborates the superiority of RBN in augmenting inter-domain transferability. This perspective transcends immediate performance metrics, offering a foundational lens through which subsequent research can more deeply understand and refine the interplay between normalization strategies and domain adaptation.

Autonomous underwater vehicle (AUV) is crucial for marine applications such as ocean data collection, pollution monitoring, and navigation. However, their limited energy resources constrain their operational duration, posing a significant challenge for long-term operations. Due to the complex and unpredictable nature of the underwater environment, AUVs allocate energy to their sensing systems to sense the surrounding environment and avoid obstacles. Existing methods focus on reducing energy consumption on AUV computing and movement, neglecting sensing energy consumption and few attempts have been made to balance the AUV energy and sensing ability with a flexible sensing system. Along these lines, we consider both AUV energy consumption and flexible sensing abilities, and propose a deep reinforcement learning-based method to Reduce Energy Consumption by AUV Sensing system (RECS). Specifically, we build an AUV sensing system in a 2-dimension space, with controllable 8-direction sensing abilities to collect the environment information dynamically. Then we divide the underwater environment into several areas and assign weights on the edges of areas based on the AUV planned path. Additionally, we dynamically switch the sensors in different directions and radii to sense the edges of the area where the AUV is located. The Artificial Potential Field (APF) method is employed to re-plan the AUV path to avoid obstacles and reach the target point effectively. Experimental results demonstrate that compared to full sensors on, our method reduces energy consumption by 53.48% and is capable of generalizing to varying environments and varying sensing system radii.

This paper shows that the semantics of programs with aggregates implemented by the solvers clingo and dlv can be characterized as extended First-Order formulas with intensional functions in the logic of Here-and-There. Furthermore, this characterization can be used to study the strong equivalence of programs with aggregates under either semantics. We also present a transformation that reduces the task of checking strong equivalence to reasoning in classical First-Order logic, which serves as a foundation for automating this procedure.

Modern Hopfield networks (MHNs) have recently gained significant attention in the field of artificial intelligence because they can store and retrieve a large set of patterns with an exponentially large memory capacity. A MHN is generally a dynamical system defined with Lagrangians of memory and feature neurons,where memories associated with in-distribution (ID) samples are represented by attractors in the feature space. One major problem in existing MHNs lies in managing out-of-distribution (OOD) samples because it was originally assumed that all samples are ID samples. To address this, we propose the rectified Lagrangian (RegLag), a new Lagrangian for memory neurons that explicitly incorporates an attractor for OOD samples in the dynamical system of MHNs. RecLag creates a trivial point attractor for any interaction matrix, enabling OOD detection by identifying samples that fall into this attractor as OOD. The interaction matrix is optimized so that the probability densities can be estimated to identify ID/OOD. We demonstrate the effectiveness of RecLag-based MHNs compared to energy-based OOD detection methods, including those using state-of-the-art Hopfield energies, across nine image datasets.

Unauthorised face recognition (FR) systems have posed significant threats to digital identity and privacy protection. To alleviate the risk of compromised identities, recent makeup transfer-based attack methods embed adversarial signals in order to confuse unauthorised FR systems. However, their major weakness is that they set up a fixed image unrelated to both the protected and the makeup reference images as the confusion identity, which in turn has a negative impact on both attack success rate and visual quality of transferred photos. In addition, the generated images cannot be recognised by authorised FR systems once attacks are triggered. To address these challenges, in this paper, we propose a Recoverable Makeup Transferred Generative Adversarial Network (RMT-GAN) which has the distinctive feature of improving its image-transfer quality by selecting a suitable transfer reference photo as the target identity. Moreover, our method offers a solution to recover the protected photos to their original counterparts that can be recognised by authorised systems. Experimental results demonstrate that our method provides significantly improved attack success rates while maintaining higher visual quality compared to state-of-the-art makeup transfer-based adversarial attack methods. Our code and supplementary materials are available on Github.

With the advancement of large-scale language modeling techniques, large multimodal models combining visual encoders with large language models have demonstrated exceptional performance in various visual tasks. Most of the current large multimodal models achieve this by mapping visual features obtained from the visual encoder into a large language model and using them as inputs alongside text for downstream tasks. Therefore, the number of visual tokens directly affects the training and inference speed of the model. There has been significant work on token pruning for visual transformers, but for large multimodal models, only relying on visual information for token pruning or compression may lead to significant loss of important information. On the other hand, the textual input in the form of a question may contain valuable information that can aid in answering the question, providing additional knowledge to the model. To address the potential oversimplification and excessive pruning that can occur with most purely visual token pruning methods, we propose a text information-guided dynamic visual token recovery mechanism that does not require training. This mechanism leverages the similarity between the question text and visual tokens to recover visually meaningful tokens with important text information while merging other less important tokens, to achieve efficient computation for large multimodal models. Experimental results demonstrate that our proposed method achieves comparable performance to the original approach while compressing the visual tokens to an average of 10\% of the original quantity.