AAAI 2025 Papers

Backdoor attacks significantly compromise the security of large language models by triggering them to output specific and controlled content. Currently, triggers for textual backdoor attacks fall into two categories: fixed-token triggers and sentence-pattern triggers. However, the former are typically easy to identify and filter, while the latter, such as syntax and style, do not apply to all original samples and may lead to semantic shifts. In this paper, inspired by cross-lingual (CL) prompts of LLMs in real-world scenarios, we propose a higher-dimensional trigger method at the paragraph level, namely CL-Attack. CL-Attack injects the backdoor by using texts with specific structures that incorporate multiple languages, thereby offering greater stealthiness and universality compared to existing backdoor attack techniques. Extensive experiments on different tasks and model architectures demonstrate that CL-Attack can achieve nearly 100 percents attack success rate with a low poisoning rate in both classification and generation tasks. We also empirically show that CL-Attack is more robust against current major defense methods compared to baseline backdoor attacks. Additionally, in response to CL-Attack, we further develop a new defense called TranslateDefense, which can partially mitigate the impact of CL-Attack.

Tool learning enables Large Language Models (LLMs) to interact with the external environment by invoking tools, enriching the accuracy and capability scope of LLMs. However, previous works predominantly focus on improving the model's tool-utilizing accuracy and the ability to generalize to new, unseen tools, excessively forcing LLMs to adjust specific tool-invoking pattern without considering the harm to the model's general performance. This deviates from the actual applications and original intention of integrating tools to enhance the model. To tackle this problem, we dissect the capability trade-offs by examining the hidden representation changes and the gradient-based importance score of the model's components. Based on the analysis result, we propose a Component Importance-based Tool-utilizing ability Injection method (CITI). According to the gradient-based importance score of different components, it alleviates the capability conflicts caused by the fine-tuning process by applying distinct training strategies to different components. CITI applies Mixture-Of-LoRA (MOLoRA) for important components. Meanwhile, it fine-tunes the parameters of a few components deemed less important in the backbone of the LLM, while keeping other parameters frozen. CITI can effectively enhance the model's tool-utilizing capability without excessively compromising its general performance. Experimental results demonstrate that our approach achieves outstanding performance across a range of evaluation metrics.

Data sharing is necessary for AI to be widely used, but sharing sensitive data with others with privacy is risky. To solve these problems, it is necessary to synthesize realistic tabular data. In many cases, tabular data contains a mixture of continuous, mixed, categorical columns. Moreover, columns of the same type may have multimodal distribution or be highly imbalanced. These issues make it challenging to synthesize tabular data. The synthesized tabular data should reflect the relational meaning between columns of tabular data, so modeling the probability distribution of tabular data is a non-trivial task. Traditional tabular data synthesizing models are based on GAN or diffusion models and are built using fully connected or convolutional layers. However, fully connected layers have the disadvantage of low inductive bias, and convolutional layers are not invariant to the column order of tabular data. Therefore, we assume that converting tabular data into graph-structured data and using a graph neural network would produce better synthetic data than using fully connected layers or convolutional layers. Our study aims to show that GANs constructed with graph neural networks can outperform existing GAN models using fully connected layers or convolutional layers. We propose CG-TGAN, a conditional GAN built using graph neural networks. To learn how to synthesize realistic data, the graph neural networks in the discriminator and generator learn graph-level tasks and node-level tasks together. The discriminator of CG-TGAN learns a graph-level task to distinguish between real and synthetic data and node-level tasks to predict the value of the target node. CG-TGAN’s generator learns a graph-level task to synthesize an overall graph similar to real data and node-level tasks to learn how to synthesize a fake graph with the proper relation between nodes. In this paper, we show that CG-TGAN outperforms GAN-based models and is comparable to diffusion-based models.

When dealing with multi-view data, the heterogeneity of data attributes across different views often leads to label ambiguity. To effectively address this challenge, this paper designs a Multi-View Partial-Label Learning (MVPLL) framework, where each training instance is described by multiple view features and associated with a set of candidate labels, among which only one is correct. The key to deal with such problem lies in how to effectively fuse multi-view information and accurately disambiguate these ambiguous labels. In this paper, we propose a novel approach named CFDM, which explores the consistency and complementarity of multi-view data by multi-view contrastive fusion and reduces label ambiguity by multi-class contrastive prototype disambiguation. Specifically, we first extract view-specific representations using multiple view-specific autoencoders, and then integrate multi-view information through both inter-view and intra-view contrastive fusion to enhance the distinctiveness of these representations. Afterwards, we utilize these distinctive representations to establish and update prototype vectors for each class within each view. Based on these, we apply contrastive prototype disambiguation to learn global class prototypes and accordingly reduce label ambiguity. In our model, multi-view contrastive fusion and multi-class contrastive prototype disambiguation are conducted mutually to enhance each other within a coherent framework, leading to a more ideal classification performance. Experimental results on multiple datasets have demonstrated that our proposed method is superior to other state-of-the-art methods.

Large Language Models (LLMs) have recently achieved impressive results in complex reasoning tasks through Chain of Thought (CoT) prompting. However, most existing CoT methods rely on using the same prompts, whether manually designed or automatically generated, to handle the entire dataset. This one-size-fits-all approach may fail to meet the specific needs arising from the diversities within a single dataset. To solve this problem, we propose the Clustered Distance-Weighted Chain of Thought (CDW-CoT) method, which dynamically constructs prompts tailored to the characteristics of each data instance by integrating clustering and prompt optimization techniques. Our method employs clustering algorithms to categorize the dataset into distinct groups, from which a candidate pool of prompts is selected to reflect the inherent diversity within the dataset. For each cluster, CDW-CoT trains the optimal prompt probability distribution tailored to their specific characteristics. Finally, it dynamically constructs a unique prompt probability distribution for each test instance, based on its proximity to cluster centers, from which prompts are selected for reasoning. CDW-CoT consistently outperforms traditional CoT methods across six datasets, including commonsense, symbolic, and mathematical reasoning tasks. Specifically, when compared to manual CoT, CDW-CoT achieves an average accuracy improvement of 25.34% on LLaMA2 (13B) and 15.72% on LLaMA3 (8B).

Video Moment Retrieval (VMR) involves locating specific moments within a video based on natural language queries. However, existing VMR methods that employ various strategies for cross-modal alignment still face challenges such as limited understanding of fine-grained semantics, semantic overlap, and sparse constraints. To address these limitations, we propose a novel Concept Decomposition Transformer (CDTR) model for VMR. CDTR introduces a semantic concept decomposition module that disentangles video moments and sentence queries into concept representations, reflecting the relevance between various concepts and capturing fine-grained semantics which is crucial for cross-modal matching. These decomposed concept representations are then used as pseudo-labels, determined as positive or negative samples by adaptive concept-specific thresholds. Subsequently, fine-grained concept alignment is performed in video intra-modal and textual-visual cross-modal, aligning different conceptual components within features, enhancing the model's ability to distinguish fine-grained semantics, and alleviating issues related to semantic overlap and sparse constraints. Comprehensive experiments demonstrate the effectiveness of the CDTR, outperforming state-of-the-art methods on three widely used datasets: QVHighlights, Charades-STA, and TACoS.

Unsupervised Person Re-identification (Re-ID) aims to identify the same person shot from non-overlapping cameras without any annotated data. In this task, attributes such as contrast, saturation, and resolution of the camera cause the deviation in target features. Since the camera label is readily available, they are employed to achieve the constraints across cameras and smooth the deviations during the model training phase. However, features from the same camera are prone to generating false positives due to the identical camera properties, which induce camera deviations on pseudo-label assignment. To address this problem, this paper proposes a novel camera-unbiased method named Camera Deviation Elimination Learning (CDE-Learning). In CDE-Learning, the Camera Deviation Compensation (CDC) module is designed to align data distributions from disparate cameras to decouple camera information from identity information during the pseudo-label allocation. Our Camera Deviation Balancing (CDB) module integrates different camera constraints in a united loss and adjusts camera constraints by constructing contrastive pairs between intra-camera and inter-camera. After explicit constraints, the Camera Attribution Auxiliary (CAA) task predicts whether a pair of images originates from the same camera to implicitly enhance the capacity to distinguish the camera deviation. We demonstrated the superior performance of the proposed CDE-Learning on benchmark datasets.

Unsupervised domain adaptation (UDA) refers to a domain adaptation framework in which a learning model is trained based on the labeled samples on the source domain and unlabelled ones in the target domain. The dominant existing methods in the field that rely on the classical covariate shift assumption to learn domain-invariant feature representation have yielded suboptimal performance under label distribution shift. In this paper, we propose a novel Conditional Adversarial SUpport ALignment (CASUAL) whose aim is to minimize the conditional symmetric support divergence between the source’s and target domain’s feature representation distributions, aiming at a more discriminative representation for the classification task. We also introduce a novel theoretical target risk bound, which justifies the merits of aligning the supports of conditional feature distributions compared to the existing marginal support alignment approach in the UDA settings. We then provide a complete training process for learning in which the objective optimization functions are precisely based on the proposed target risk bound. Our empirical results demonstrate that CASUAL outperforms other state-of-the-art methods on different UDA benchmark tasks under different label shift conditions.

Recently, prompt tuning methods for pre-trained models have demonstrated promising performance in Class Incremental Learning (CIL). These methods typically involve learning task-specific prompts and predicting the task ID to select the appropriate prompts for inference. However, inaccurate task ID predictions can cause severe inconsistencies between the prompts used during training and inference, leading to knowledge forgetting and performance degradation. Additionally, existing prompt tuning methods rely solely on the pre-trained model to predict task IDs, without fully leveraging the knowledge embedded in the learned prompt parameters, resulting in inferior prediction performance. To address these issues, we propose a novel Cyclic Prompt Aggregation (CAPrompt) method that eliminates the dependency on task ID prediction by cyclically aggregating the knowledge from different prompts. Specifically, rather than predicting task IDs, we introduce an innovative prompt aggregation strategy during both training and inference to overcome prompt inconsistency by utilizing a weighted sum of different prompts. Thorough theoretical analysis demonstrates that under concave conditions, the aggregated prompt achieves lower error compared to selecting a single task-specific prompt. Consequently, we incorporate a concave constraint and a linear constraint to guide prompt learning, ensuring compliance with the concave condition requirement. Furthermore, to fully exploit the prompts and achieve more accurate prompt weights, we develop a cyclic weight prediction strategy. This strategy begins with equal weights for each task and automatically adjusts them to more appropriate values in a cyclical manner. Experiments on various datasets demonstrate that our proposed CAPrompt outperforms state-of-the-art methods by 2%-3%.

Existing learning-based stereo image codec adopt sophisticated transformation with simple entropy models derived from single image codecs to encode latent representations. However, those entropy models struggle to effectively capture the spatial-disparity characteristics inherent in stereo images, which leads to suboptimal rate-distortion results. In this paper, we propose a stereo image compression framework, named CAMSIC. CAMSIC independently transforms each image to latent representation and employs a powerful decoder-free Transformer entropy model to capture both spatial and disparity dependencies, by introducing a novel content-aware masked image modeling (MIM) technique. Our content-aware MIM facilitates efficient bidirectional interaction between prior information and estimated tokens, which naturally obviates the need for an extra Transformer decoder. Experiments show that our stereo image codec achieves state-of-the-art rate-distortion performance on two stereo image datasets Cityscapes and InStereo2K with fast encoding and decoding speed.

In the burgeoning domain of machine learning, the reliance on third-party services for model training and the adoption of pre-trained models have surged. However, this reliance introduces vulnerabilities to model hijacking attacks, where adversaries manipulate models to perform unintended tasks, leading to significant security and ethical concerns, like turning an ordinary image classifier into a tool for detecting faces in pornographic content, all without the model owner’s knowledge. This paper introduces Category-Agnostic Model Hijacking (CAMH), a novel model hijacking attack method capable of addressing the challenges of class number mismatch, data distribution divergence, and performance balance between the original and hijacking tasks. CAMH incorporates synchronized training layers, random noise optimization, and a dual-loop optimization approach to ensure minimal impact on the original task’s performance while effectively executing the hijacking task. We evaluate CAMH across multiple benchmark datasets and network architectures, demonstrating its potent attack effectiveness while ensuring minimal degradation in the performance of the original task.

Auditing Large Language Models (LLMs) is a crucial and challenging task. In this study, we focus on auditing black-box LLMs without access to their parameters, only to the provided service. We treat this type of auditing as a black-box optimization problem where the goal is to automatically uncover input-output pairs of the target LLMs that exhibit illegal, immoral, or unsafe behaviors. For instance, we may seek a non-toxic input that the target LLM responds to with a toxic output or an input that induces the hallucinative response from the target LLM containing politically sensitive individuals. This black-box optimization is challenging due to the scarcity of feasible points, the discrete nature of the prompt space, and the large search space. To address these challenges, we propose Curiosity-Driven Auditing for Large Language Models (CALM), which uses intrinsically motivated reinforcement learning to finetune an LLM as the auditor agent to uncover potential harmful and biased input-output pairs of the target LLM. CALM successfully identifies derogatory completions involving celebrities and uncovers inputs that elicit specific names under the black-box setting. This work offers a promising direction for auditing black-box LLMs.

Learned lossless image compression has achieved significant advancements in recent years. However, existing methods often rely on training amortized generative models on massive datasets, resulting in sub-optimal probability distribution estimation for specific testing images during encoding process. To address this challenge, we explore the connection between the Minimum Description Length (MDL) principle and Parameter-Efficient Transfer Learning (PETL), leading to the development of a novel content-adaptive approach for learned lossless image compression, dubbed CALLIC. Specifically, we first propose a content-aware autoregressive self-attention mechanism by leveraging convolutional gating operations, termed Masked Gated ConvFormer (MGCF), and pretrain MGCF on training dataset. Cache then Crop Inference (CCI) is proposed to accelerate the coding process. During encoding, we decompose pretrained layers, including depth-wise convolutions, using low-rank matrices and then adapt the incremental weights on testing image by Rate-guided Progressive Fine-Tuning (RPFT). RPFT fine-tunes with gradually increasing patches that are sorted in descending order by estimated entropy, optimizing learning process and reducing adaptation time. Extensive experiments across diverse datasets demonstrate that CALLIC sets a new state-of-the-art (SOTA) for learned lossless image compression.

Deep learning (e.g., Transformer) has been widely and successfully used in multivariate time series forecasting (MTSF). Unlike existing methods that focus on training models from a single modal of time series input, large language models (LLMs) based MTSF methods with cross-modal text and time series input have recently shown great superiority, especially with limited temporal data. However, current LLM-based MTSF methods usually focus on adapting and fine-tuning LLMs, while neglecting the distribution discrepancy between textual and temporal input tokens, thus leading to sub-optimal performance. To address this issue, we propose a novel Cross-Modal LLM Fine-Tuning (CALF) framework for MTSF by reducing the distribution discrepancy between textual and temporal data, which mainly consists of the temporal target branch with temporal input and the textual source branch with aligned textual input. To reduce the distribution discrepancy, we develop the cross-modal match module to first align cross-modal input distributions. Additionally, to minimize the modality distribution gap in both feature and output spaces, feature regularization loss is developed to align the intermediate features between the two branches for better weight updates, while output consistency loss is introduced to allow the output representations of both branches to correspond effectively. Thanks to the modality alignment, CALF establishes state-of-the-art performance for both long-term and short-term forecasting tasks with low computational complexity, and exhibits favorable few-shot and zero-shot abilities similar to that in LLMs.

Open vocabulary object detection (OVOD) task aims to detect objects of novel categories beyond the base categories in the training set. To this end, the detector needs to access image-text pairs containing rich semantic information or the visual language pre-trained model (VLM) learned on them. Recent OVOD methods rely on knowledge distillation from VLMs. However, there are two main problems in current methods: (1) Current knowledge distillation frameworks fail to take advantage of the global category information of VLMs and thus fail to learn category-specific knowledge. (2) Due to the overfitting phenomenon of base categories during training, current OVOD networks generally have the problem of suppressing novel categories as background. To address these two problems, we propose a Category Aware Knowledge Extraction framework (CAKE), which consists of a Category-Specific Knowledge Distillation branch (CSKD) and a Category Generalization Region Proposal Network (CG-RPN). CSKD can more fully extract category-strong related information through category-specific distillation, and it is also conducive to filtering the exclusion problem between individuals of the same category; in this process, the model constructs a category-specific feature set to maintain high-quality category features. CG-RPN leverages the guidance of feature set to adjust the confidence scores of region proposals, thereby mining proposals that potentially contain novel categories of objects. Extensive experiments show that our method can plug and play well with many existing methods and significantly improve their detection performance. Moreover, our CAKE framework can reach the-state-of-the-art performance on OV-COCO and OV-LVIS datasets.

The field of video generation has expanded significantly in recent years, with controllable and compositional video generation garnering considerable interest. Most methods rely on leveraging annotations such as text, objects' bounding boxes, and motion cues, which require substantial human effort and thus limit their scalability. In contrast, we address the challenge of controllable and compositional video generation without any annotations by introducing a novel unsupervised approach. Our model is trained from scratch on a dataset of unannotated videos. At inference time, it can compose plausible novel scenes and animate objects by placing object parts at the desired locations in space and time. The core innovation of our method lies in the unified control format and the training process, where video generation is conditioned on a randomly selected subset of pre-trained self-supervised local features. This conditioning compels the model to learn how to inpaint the missing information in the video both spatially and temporally, thereby learning the inherent compositionality of a scene and the dynamics of moving objects. The abstraction level and the imposed invariance of the conditioning input to minor visual perturbations enable control over object motion by simply using the same features at all the desired future locations. We call our model CAGE, which stands for visual Composition and Animation for video GEneration. We conduct extensive experiments to validate the effectiveness of CAGE across various scenarios, demonstrating its capability to accurately follow the control and to generate high-quality videos that exhibit coherent scene composition and realistic animation.

Computer-aided design (CAD) significantly enhances the efficiency, accuracy, and innovation of design processes by enabling precise 2D and 3D modeling, extensive analysis, and optimization. Existing methods for creating CAD models rely on latent vectors or point clouds, which are difficult to obtain, and storage costs are substantial. Recent advances in Multimodal Large Language Models (MLLMs) have inspired researchers to use natural language instructions and images for CAD model construction. However, these models still struggle with inferring accurate 3D spatial location and orientation, leading to inaccuracies in determining the spatial 3D starting points and extrusion directions for constructing geometries. This work introduces CAD-GPT, a CAD synthesis method with spatial reasoning-enhanced MLLM that takes either a single image or a textual description as input. To achieve precise spatial inference, our approach introduces a 3D Modeling Spatial Mechanism. This method maps 3D spatial positions and 3D sketch plane rotation angles into a 1D linguistic feature space using a specialized spatial unfolding mechanism, while discretizing 2D sketch coordinates into an appropriate planar space to enable precise determination of spatial starting position, sketch orientation, and 2D sketch coordinate translations. Extensive experiments demonstrate that CAD-GPT consistently outperforms existing state-of-the-art methods in CAD model synthesis, both quantitatively and qualitatively.

Cancer is a leading cause of death worldwide due to its aggressive nature and complex variability. Accurate prognosis is therefore challenging but essential for guiding personalized treatment and follow-up. Previous research often relied on single data sources, missing the opportunity to combine various types of patient information for more comprehensive survival predictions. To address these challenges, we propose a two-stage fusion method named Cross-Attention and Multimodal Low-Rank Interaction Fusion Framework (CA-MLIF). In the first stage, we propose a CA mechanism for real-time feature updates and cross-modal mutual learning to capture rich semantic information. In the second stage, we design a novel multimodal low-rank interaction fusion method for survival prediction. Specifically, we present modal attention mechanism (MAM) for feature filtration, low-rank multimodal fusion (LMF) for model complexity reduction, and optimal weight concatenation (OWC) for maximizing feature integration. Extensive experiments on two public datasets TCGA-GBMLGG and TCGA-KIRC, as well as a multi-center in-house lung adenocarcinoma (LUAD) dataset validate the effectiveness of CA-MLIF, which demonstrate that our method outperforms existing approaches in survival prediction under both pathology-gene fusion and CT-pathology fusion scenarios.

For efficient and high-fidelity local facial attribute editing, most existing editing methods either require additional fine-tuning for different editing effects or tend to affect beyond the editing regions. Alternatively, inpainting methods can edit the target image region while preserving external areas. However, current inpainting methods still suffer from the generation misalignment with facial attributes description and the loss of facial skin details. To address these challenges, (i) a novel data utilization strategy is introduced to construct datasets consisting of attribute-text-image triples from a data-driven perspective, (ii) a Causality-Aware Condition Adapter is proposed to enhance the contextual causality modeling of specific details, which encodes the skin details from the original image while preventing conflicts between these cues and textual conditions. In addition, a Skin Transition Frequency Guidance technique is introduced for the local modeling of contextual causality via sampling guidance driven by low-frequency alignment. Extensive quantitative and qualitative experiments demonstrate the effectiveness of our method in boosting both fidelity and editability for localized attribute editing. Our codes will be made publicly available.