AAAI 2025 论文列表

In this paper, we explore a novel federated multimodal instruction tuning task(FedMIT), which is significant for collaboratively fine-tuning MLLMs on different types of multimodal instruction data on distributed devices. To solve the new task, we propose a federated multimodal instruction tuning framework(Pilot). Our framework integrates two-stage of ``adapter on adapter” into the connector of the vision encoder and the LLM. In stage 1, we extract task-specific features and client-specific features from visual information. In stage 2, we build the cross-task Mixture-of-Adapters(CT-MoA) module to perform cross-task interaction. Each client can not only capture personalized information of local data and learn task-related multimodal information, but also learn general knowledge from other tasks. In addition, we introduce an adaptive parameter aggregation strategy for text training parameters, which optimizes parameter aggregation by calculating weights based on the euclidean distance between parameters, so that parameter aggregation can benefit from positive effects to the greatest extent while effectively reducing negative effects. Our framework can collaboratively exploit distributed data from different local clients to learn cross-task knowledge without being affected by the task heterogeneity during instruction tuning. The effectiveness of our method is verified in two different cross-task scenarios.

As the demand for healthier, personalized culinary experiences grows, so does the need for advanced food computation models that offer more than basic nutritional insights. However, current food computation models lack the depth to provide actionable insights like ingredient substitution or alternative cooking actions to suit users’ dietary goals. To address this, we introduce and demonstrate Pic2Prep, a multimodal conversational system that generates detailed cooking instructions, actions and ingredient lists from both images and text provided by users. The system is developed using a novel dataset generated through Stable Diffusion, where the input consists of recipe titles and ingredient lists from the Recipe1M dataset to create synthesized food images with variations. This dataset is used to fine-tune the Bootstrapping Language-Image Pre-training (BLIP) model to extract cooking instructions and ingredients from food images. Pic2Prep also employs the CookGen model, a small-scale custom generative model to derive specific cooking actions from cooking instructions. A custom mapper, trained on the Mistral model, links these actions to the corresponding ingredients, creating a comprehensive understanding of the cooking process. The system features an interactive user interface that allows users to input images and ask targeted questions, receiving real-time responses.

Physics-guided machine learning (PGML) has become a prevalent approach in studying scientific systems due to its ability to integrate scientific theories for enhancing machine learning (ML) models. However, most PGML approaches are tailored to isolated and relatively simple tasks, which limits their applicability to complex systems involving multiple interacting processes and numerous influencing features. In this paper, we propose a Physics-Guided Foundation Model (PGFM) that combines pre-trained ML models and physics-based models and leverages their complementary strengths to improve the modeling of multiple coupled processes. To effectively conduct pre-training, we construct a simulated environmental system that encompasses a wide range of influencing features and various simulated variables generated by physics-based models. The model is pre-trained in this system to adaptively select important feature interactions guided by multi-task objectives. We then fine-tune the model for each specific task using true observations, while maintaining consistency with established physical theories, such as the principles of mass and energy conservation. We demonstrate the effectiveness of this methodology in modeling water temperature and dissolved oxygen dynamics in real-world lakes. The proposed PGFM is also broadly applicable to a range of scientific fields where physics-based models are being used.

This work introduces a novel graph neural networks (GNNs)-based method to predict stream water temperature and reduce model bias across locations of different income and education levels. Traditional physics-based models often have limited accuracy because they are necessarily approximations of reality. Recently, there has been an increasing interest of using GNNs in modeling complex water dynamics in stream networks. Despite their promise in improving the accuracy, GNNs can bring additional model bias through the aggregation process, where node features are updated by aggregating neighboring nodes. The bias can be especially pronounced when nodes with similar sensitive attributes are frequently connected. We introduce a new method that leverages physical knowledge to represent the node influence in GNNs, and then utilizes physics-based influence to refine the selection and weights over the neighbors. The objective is to facilitate equitable treatment over different sensitive groups in the graph aggregation, which helps reduce spatial bias over locations, especially for those in underprivileged groups. The results on the Delaware River Basin demonstrate the effectiveness of the proposed method in preserving equitable performance across locations in different sensitive groups.

Neural Radiance Fields (NeRF) has achieved remarkable success in synthesizing impressive novel views. However, existing methods usually fail to handle scenes with adverse lighting conditions caused by external time variations and different camera settings, leading to poor visual quality. To address this challenge, we propose a physical-aware NeRF for efficient exposure correction, named PHY-NeRF. Specifically, we design Adaptive Lighting Particles inspired by the theory of light scattering and absorption, which can adjust the illumination intensity during volume rendering. Subsequently, we can handle scenes with different lighting conditions by jointly optimizing camera parameters and these lighting particles. Moreover, to promote natural brightness transitions, we devise a global illumination consistency module to control the lighting intensity across views at the feature level while completing more details. Benefiting from the above designs, our PHY-NeRF can tackle arbitrary low-light or overexposed scenes in an unsupervised manner. Extensive experiments show that our PHY-NeRF achieves state-of-the-art results in addressing adverse lighting problems while ensuring high rendering efficiency.

Embedding links in brand logos is a promising technology, which allows consumers to access the online information of products by capturing physical logo images. Previous physical data hiding methods primarily embed data within cover media in a global manner, making them ineffective for processing brand logos in vector graphics format with a transparent background. To address this issue, we propose in this paper a novel physical deep hiding scheme for invisibly embedding links in printed trademarks. Specifically, the encoder embeds links only into the area of the brand logo under the constraints of a mask, which is generated from the transparency information of the logo image. A background variation distortion is introduced into the distortion layer that approximate practical logo print-camera environments, such that the decoder could be learnt to retrieve the link from the camera-captured logo with various backgrounds. A feature prompt subspace modulator is further proposed and employed in the encoder to enhance the invisibility of the encoded logo pattern and in the decoder to boost hyperlink extraction accuracy. Various experiments have been conducted to demonstrate the advantage of our proposed method for embedding links in printed brand logos, which provides reliable extraction accuracy under both simulated and real scenarios.

Recent works on remote PhotoPlethysmoGraphy (rPPG) estimation typically use techniques like CNNs and Transformers to encode implicit features from facial videos for prediction. These methods learn to directly map facial videos to the static values of rPPG signals, overlooking the inherent dynamic characteristics of rPPG sequence. Moreover, the rPPG signal is extremely weak and highly susceptible to interference from various sources of noise, including illumination conditions, head movements, and variations in skin tone. To address these limitations, we propose a Physiology-based dynamicity disentangled diffusion (PhysDiff) model particularly designed for robust rPPG estimation. PhysDiff leverages the diffusion model to learn the distribution of quasi-periodic rPPG signal and uses a dynamicity disentanglement strategy to capture two dynamic characteristics in temporal rPPG signal, i.e., trend and amplitude. This disentanglement is motivated by the underlying dynamic physiological processes of vasodilation and vasoconstriction, ensuring a more precise representation of the rPPG signal. The disentangled components are then used as pivotal conditions in the proposed spatial-temporal hybrid denoiser for rPPG reconstruction. Besides, we introduce a periodicity-based multi-hypothesis selection strategy in model inference, which compares the natural periodicity of multiple generated rPPG hypotheses and selects the most favorable one as the final prediction. Extensive experiments on four datasets demonstrate that our PhysDiff significantly outperforms prior methods on both intra-dataset and cross-dataset testing.

Single-Domain Generalized Object Detection (S-DGOD) aims to train on a single source domain for robust performance across a variety of unseen target domains by taking advantage of an object detector. Existing S-DGOD approaches often rely on data augmentation strategies, including a composition of visual transformations, to enhance the detector's generalization ability. However, the absence of real-world prior knowledge hinders data augmentation from contributing to the diversity of training data distributions. To address this issue, we propose PhysAug, a novel physical model-based non-ideal imaging condition data augmentation method, to enhance the adaptability of the S-DGOD tasks. Drawing upon the principles of atmospheric optics, we develop a universal perturbation model that serves as the foundation for our proposed PhysAug. Given that visual perturbations typically arise from the interaction of light with atmospheric particles, the image frequency spectrum is harnessed to simulate real-world variations during training. This approach fosters the detector to learn domain-invariant representations, thereby enhancing its ability to generalize across various settings. Without altering the network architecture or loss function, our approach significantly outperforms the state-of-the-art across various S-DGOD datasets. In particular, it achieves a substantial improvement of 7.3% and 7.2% over the baseline on DWD and Cityscape-C, highlighting its enhanced generalizability in real-world settings.

Deep neural networks (DNNs) have achieved remarkable success in widespread applications. Meanwhile, its vulnerability towards carefully crafted adversarial attacks captures special attention. Not only adversarial perturbations in digital space will fool the target DNNs-based detectors making a wrong decision, but also actually printed patches can be camouflaged to defeat detectors in physical space. In particular, multi-view physical adversarial attacks pose a more serious threat to practical scenarios. The existing attacks are still challenged in three aspects, i.e., high-cost data augmentation, attack performance gap between digital and physical space, and low attack transferability across DNNs. To overcome the challenges, we introduce PhyCamo, a robust physical camouflage framework based on contrastive learning that distinguishes itself from prior research in various critical ways: (1) data augmentation - it utilizes the diffusion model for data augmentation to efficiently simulate sophisticated physical dynamics in real-world; (2) robustness - it leverages contrastive learning to optimize physical camouflage against encoders with the state-of-the-art (SOTA) attack performance; (3) transferability - it mitigates the model-specific noise in the optimization by adopting diverse input methods, thereby amplifying the transferability between models. Extensive experiments are carried out on a car dataset, a tank dataset, and a pedestrian dataset, comparing with 6 classic multi-view physical adversarial attacks in both digital and physical spaces. The results demonstrate PhyCamo’s superior performance. For instance, it generates more effective physical camouflage (with higher attack success rate~×1.26 and reduce the model's average precision by 55%). PhyCamo can also help to improve the robustness of detectors through adversarial training, which contributes to the application of deep neural networks in the field of security sensitivity.

Recent advancements in text-to-speech and speech conversion technologies have enabled the creation of highly convincing synthetic speech. While these innovations offer numerous practical benefits, they also cause significant security challenges when maliciously misused. Therefore, there is an urgent need to detect these synthetic speech signals. Phoneme features provide a powerful speech representation for deepfake detection. However, previous phoneme-based detection approaches typically focused on specific phonemes, overlooking temporal inconsistencies across the entire phoneme sequence. In this paper, we develop a new mechanism for detecting speech deepfakes by identifying the inconsistencies of phoneme-level speech features. We design an adaptive phoneme pooling technique that extracts sample-specific phoneme-level features from frame-level speech data. By applying this technique to features extracted by pre-trained audio models on previously unseen deepfake datasets, we demonstrate that deepfake samples often exhibit phoneme-level inconsistencies when compared to genuine speech. To further enhance detection accuracy, we propose a deepfake detector that uses a graph attention network to model the temporal dependencies of phoneme-level features. Additionally, we introduce a random phoneme substitution augmentation technique to increase feature diversity during training. Extensive experiments on four benchmark datasets demonstrate the superior performance of our method over existing state-of-the-art detection methods.

Phishing attacks are a major threat to online security, exploiting user vulnerabilities to steal sensitive information. Various methods have been developed to counteract phishing, each with varying levels of accuracy, but they also face notable limitations. In this study, we introduce PhishAgent, a multimodal agent that combines a wide range of tools, integrating both online and offline knowledge bases with Multimodal Large Language Models (MLLMs). This combination leads to broader brand coverage, which enhances brand recognition and recall. Furthermore, we propose a multimodal information retrieval framework designed to extract the relevant top k items from offline knowledge bases, using available information from a webpage, including logos and HTML. Our empirical results, based on three real-world datasets, demonstrate that the proposed framework significantly enhances detection accuracy and reduces both false positives and false negatives, while maintaining model efficiency. Additionally, PhishAgent shows strong resilience against various types of adversarial attacks.

Single Domain Generalization (SDG) is critical in medical imaging applications. Recently, Vision Foundation Models (VFMs) have spearheaded a trend in AI development due to their robust generalizability and versatility. This work aims to fully explore the generalization capabilities of VFMs alongside the domain-specific expertise of specialized models, thoroughly investigating the boundaries of their respective capabilities, thereby collaboratively addressing SDG challenges within medical imaging. We propose a framework for Collaborative reasoning between Specialized and Universal models for Single Domain Generalization (CollaSU-SDG) in medical imaging. Specifically, we first design a model-aware perturbation injection method from the perspective of single-source domain data, enabling differentiated and adaptive perturbation injection for two different scales of models. Then, a domain expansion adapter is designed for the VFM to adapt to the augmented single-source domain medical data. Lastly, we introduce an adaptive hierarchical transfer and dynamic dense prompting method that facilitate collaborative reasoning between the specialized and universal models, eliminating the need for explicit prompts. Through these designs, CollaSU-SDG fully leverages the strengths of both specialized and universal models, achieving robust out-of-distribution generalization capabilities on single-source domain data. Experimental results demonstrate that CollaSU-SDG significantly advances the state-of-the-art performance across a wide range of medical datasets. All the code will be publicly available.

Persuasion is important in numerous situations like healthy habit promotion, and emotional support. As AI gets more involved in our daily life, it becomes critical to study how they can persuade humans and how persuasive they are. In this talk, I will cover (1) how to build such persuasive AI systems that can persuade, negotiate, and cooperate with other humans in the game of Diplomacy. (2) I will also discuss how humans perceive such specialized AI systems. This study validates the necessity of California's Autobot Law and proposes guidance to regulate such systems. (3) As these systems become more powerful, AI safety problems become more important. So I will describe how to persuade AI models to jailbreak them and study AI safety problems. Finally, I will conclude with my long-term vision to further study persuasion from a multi-angle approach that combines Artificial Intelligence, Human-Computer Interaction, and social sciences.

Correctly assessing the malignancy of breast lesions identified during ultrasound examinations is crucial for effective clinical decision-making. However, the current "gold standard" relies on manual BI-RADS scoring by clinicians, often leading to unnecessary biopsies and a significant mental health burden on patients and their families. In this paper, we introduce PersonalizedUS, an interpretable machine learning system that leverages recent advances in conformal prediction to provide precise and personalized risk estimates with local coverage guarantees and sensitivity, specificity, and predictive values above 0.9 across various threshold levels. In particular, we identify meaningful lesion subgroups where distribution-free, model-agnostic conditional coverage holds, with approximately 90% of our prediction sets containing only the ground truth in most lesion subgroups, thus explicitly characterizing for which patients the model is most suitably applied. Moreover, we make available a curated tabular dataset of 1936 biopsied breast lesions from a recent observational multicenter study and benchmark the performance of several state-of-the-art learning algorithms. We also report a successful case study of the deployed system in the same multicenter context. Concrete clinical benefits include up to a 65% reduction in requested biopsies among BI-RADS 4a and 4b lesions, with minimal to no missed cancer cases.

Sleep staging is important for monitoring sleep quality and diagnosing sleep-related disorders. Recently, numerous deep learning-based models have been proposed for automatic sleep staging using polysomnography recordings. Most of them are trained and tested on the same labeled datasets which results in poor generalization to unseen target domains. However, they regard the subjects in the target domains as a whole and overlook the individual discrepancies, which limits the model's generalization ability to new patients (i.e., unseen subjects) and plug-and-play applicability in clinics. To address this, we propose a novel Source-Free Unsupervised Individual Domain Adaptation (SF-UIDA) framework for sleep staging, leveraging sequential cross-view contrasting and pseudo-label based fine-tuning. It is actually a two-step subject-specific adaptation scheme, which enables the source model to effectively adapt to newly appeared unlabeled individual without access to the source data. It meets the practical needs in real-world scenarios, where the personalized customization can be plug-and-play applied to new ones. Our framework is applied to three classic sleep staging models and evaluated on three public sleep datasets, achieving the state-of-the-art performance.

Lip reading aims to predict spoken language by analyzing lip movements. Despite advancements in lip reading technologies, performance degrades when models are applied to unseen speakers due to their sensitivity to variations in visual information such as lip appearances. To address this challenge, speaker adaptive lip reading technologies have advanced by focusing on effectively adapting a lip reading model to target speakers in the visual modality. However, the effectiveness of adapting language information, such as vocabulary choice, of the target speaker has not been explored in previous works. Additionally, existing datasets for speaker adaptation have limited vocabulary sizes and pose variations, which restrict the validation of previous speaker-adaptive methods in real-world scenarios. To address these issues, we propose a novel speaker-adaptive lip reading method that adapts a pre-trained model to target speakers at both vision and language levels. Specifically, we integrate prompt tuning and the LoRA approach, applying them to a pre-trained lip reading model to effectively adapt the model to target speakers. Furthermore, to validate its effectiveness in real-world scenarios, we introduce a new dataset, VoxLRS-SA, derived from VoxCeleb2 and LRS3. It contains a vocabulary of approximately 100K words, offers diverse pose variations, and enables the validation of adaptation methods in the wild, sentence-level lip reading for the first time in English. Through various experiments, we demonstrate that the existing speaker-adaptive method also improves performance in the wild at the sentence level. Moreover, we show that the proposed method achieves larger improvements compared to the previous works.

The lack of personalization in early education can often leave students with weak foundational skills, causing said students to be behind in their studies. Personalized learning, the idea of tailoring a unique lesson plan to a student, has been shown to improve the understanding of content learned. Robots utilizing personalization techniques in educational settings, coined social robots, have been able to form a connection with students, thereby keeping them engaged while learning. This proposal seeks to study the effects of AI-driven social robotic tutors coupled with personalized learning on early childhood education. The study will consist of five groups of K-4 students: two groups learning while utilizing both a social robot and a tablet (one group with personalized learning and the other without), the two groups interacting with only the tablet (with and without personalization), and the last group learning utilizing both a non-personalized learning tablet and a non-social robot. This study aims to determine whether the combination of robotic interaction and personalized learning leads to better outcomes than solely tablet-based or non-personalized methods. This study will focus on teaching mathematics to the participants. Pre and post-tests will measure learning progress, and the influence of robot interaction on student engagement will also be evaluated. It is expected that the students with social robotic tutors and personalized learning tablets will show the greatest knowledge retention, outperforming all other categories. These findings could have significant implications for the integration of AI and robotics in early education, potentially revolutionizing how personalized learning is implemented therefore improving educational outcomes for young learners.

Federated Transfer Learning (FTL) is a popular approach to solve the problem of heterogeneous feature space and label distribution. Among the mainstream strategies for FTL, parameter decoupling, which balance the impact of a single global model and multiple personalized models under data heterogeneity, has attracted the attention of many researchers. However, few attacks have been proposed to evaluate the privacy risk of FTL. We find that the fine-tuned structures and the gradient update mechanisms of parameter decoupling would be more likely to leak personalized information for the server to infer private labels. Based on our findings, we propose the label inference attack that combines meta classifier with contrastive learning in FTL. Our experiments show that the proposed attack has ability to extract local personalized information from the differences before and after fine-tuning to improve the accuracy of the attack in the absence of a downstream model. Our research can reveal potential privacy risks in FTL and motivate more research on private and secure FTL.

The existing federated learning (FL) methods for spatio-temporal forecasting fail to capture the inherent spatio-temporal heterogeneity, which calls for personalized FL (PFL) methods to model the spatio-temporally variant representations. While contrastive learning is promising in tackling spatio-temporal heterogeneity, the existing methods are noneffective in distinguishing positive and negative pairs and can hardly apply to PFL paradigm. To tackle this limitation, we propose a novel PFL method, named Federated dUal sEmantic aLignment-based contraStive learning (FUELS), which can adaptively align positive and negative pairs based on semantic similarity, thereby injecting precise spatio-temporal heterogeneity into the latent representation space by auxiliary contrastive tasks. From temporal perspective, a hard negative filtering module is introduced to dynamically align heterogeneous temporal representations for the supplemented intra-client contrastive task. From spatial perspective, we design lightweight-but-efficient prototypes as client-level semantic representations, based on which the server evaluates spatial similarity and yields client-customized global prototypes for the supplemented inter-client contrastive task. Extensive experiments demonstrate that FUELS outperforms state-of-the-art methods, with impressive communication cost reduction.