United States

Recent studies have highlighted the fragility of safety alignment in large language models (LLMs), revealing that fine-tuning on task-specific datasets mixing even a few harmful instructions can significantly compromise their safety. Existing methods to counteract fine-tuning attacks typically require substantial computational resources. Even with parameter-efficient techniques like LoRA, gradient updates remain essential. To address these challenges, we propose $N$euron-$L$evel $S$afety $R$ealignment ($NLSR$), a training-free framework that restores the safety of LLMs based on the similarity difference of safety-critical neurons before and after fine-tuning. The core of our framework is first to construct a safety reference model from an initially aligned model to amplify safety-related features in neurons. We then utilize this reference model to identify safety-critical neurons, which we prepare as patches. Finally, we selectively restore only those neurons that exhibit significant similarity differences by transplanting these prepared patches, thereby minimally altering the fine-tuned model. Extensive experiments demonstrate significant safety enhancements in fine-tuned models across a range of downstream tasks, while greatly maintaining task-level accuracy. Our findings suggest regions of some safety-critical neurons show noticeable differences after fine-tuning, which can be effectively corrected by transplanting neurons from the reference model without requiring additional training.

AAAI 2025

NLSR: Neuron-Level Safety Realignment of Large Language Models Against Harmful Fine-Tuning

safety and robustness

snlp

poster

We are pleased to announce the Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI-25), which will be held in Philadelphia, Pennsylvania at the Pennsylvania Convention Center from February 25 to March 4, 2025.

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-25 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.

### [Invited Speakers](https://aaai.org/conference/aaai/aaai-25/aaai-25-invited-speakers/)

Register [here](https://aaai.org/conference/aaai/aaai-25/registration/)

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-25 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.



The self-attention mechanism has been adopted in various popular message passing neural networks (MPNNs), enabling the model to adaptively control the amount of information that flows along the edges of the underlying graph. Such attention-based MPNNs (Att-GNNs) have also been used as a baseline for multiple studies on explainable AI (XAI) since attention has steadily been seen as natural model interpretations, while being a viewpoint that has already been popularized in other domains (e.g., natural language processing and computer vision). However, existing studies often use naive calculations to derive attribution scores from attention, undermining the potential of attention as interpretations for Att-GNNs. In our study, we aim to fill the gap between the widespread usage of Att-GNNs and their potential explainability via attention. To this end, we propose GAtt, edge attribution calculation method for self-attention MPNNs based on the computation tree, a rooted tree that reflects the computation process of the underlying model. Despite its simplicity, we empirically demonstrate the effectiveness of GAtt in three aspects of model explanation: faithfulness, explanation accuracy, and case studies by using both synthetic and real-world benchmark datasets. In all cases, the results demonstrate that GAtt greatly improves edge attribution scores, especially compared to the previous naive approach. We intend to make our source code publicly available after acceptance.

Faithful and Accurate Self-Attention Attribution for Message Passing Neural Networks via the Computation Tree Viewpoint

Tabular data poses unique challenges due to its heterogeneous nature, combining both continuous and categorical variables. Existing approaches often struggle to effectively capture the underlying structure and relationships within such data. We propose GFTab (Geodesic Flow Kernels for Semi-Supervised Learning on Mixed-Variable Tabular Dataset), a semi-supervised framework specifically designed for tabular datasets. GFTab incorporates three key innovations: 1) Variable-specific corruption methods tailored to the distinct properties of continuous and categorical variables, 2) A Geodesic flow kernel based similarity measure to capture geometric changes between corrupted inputs, and 3) Tree-based embedding to leverage hierarchical relationships from available labeled data. To rigorously evaluate GFTab, we curate a comprehensive set of 21 tabular datasets spanning various domains, sizes, and variable compositions. Our experimental results show that GFTab outperforms existing ML/DL models across many of these datasets, particularly in settings with limited labeled data.

Geodesic Flow Kernels for Semi-Supervised Learning on Mixed-Variable Tabular Dataset

Unmanned Aerial Vehicle Object Detection (UAVOD) presents unique challenges due to varying altitudes, dynamic backgrounds, and the small size of objects. Traditional detection methods often struggle with these challenges, as they typically rely on visual feature only and fail to extract the semantic relations between the objects. To address these limitations, we propose a novel approach named Self-Prompting Analogical Reasoning (SPAR). Our method utilizes the vision-language model (CLIP) to generate context-aware prompts based on image feature, providing rich semantic information that guides analogical reasoning. SPAR includes two main modules: self-prompting and analogical reasoning. Self-prompting module based on learnable description and CLIP-text encoder generates context-aware prompt by combining specific image feature; then an objectness prompt score map is produced by computing the similarity between pixel-level features and context-aware prompt. With this score map, multi-scale image features are enhanced and pixel-level features are chosen for graph construction. While for analogical reasoning module, graph nodes consists of category-level prompt nodes and pixel-level image feature nodes. Analogical inference is based graph convolution. Under the guidance of category-level nodes, different-scale object features  have been enhanced, which helps achieve more accurate detection of challenging objects. Extensive experiments illustrate  that SPAR outperforms traditional methods, offering a more robust and accurate solution for UAVOD.

Self-Prompting Analogical Reasoning for UAV Object Detection

The infrequent occurrence of overfitting in deep neural networks is perplexing: seemingly at odds with theoretical predictions, expanding models typically only enhances performance in practical applications. However, what if overfitting does occur, albeit confined to specific sub-regions of the data space? Here, we introduce a novel score that captures the forgetting rate of deep models on validation data. We posit that this score quantifies local overfitting: a decline in performance confined to certain regions of the data space. We then show empirically that local overfitting occurs regardless of the presence of traditional overfitting. Using the framework of deep over-parametrized linear models, we offer a certain theoretical characterization of forgotten knowledge, and show that it correlates with knowledge forgotten by real deep models. Finally, we devise a new ensemble method that aims to recover forgotten knowledge, relying solely on the training history of a single network. This method enhances the performance of any trained model without incurring additional training costs. Extensive empirical evaluations demonstrate the efficacy of our method across multiple datasets, contemporary neural network architectures, and training protocols.

On Local Overfitting and Forgetting in Deep Neural Networks

Language steganography in social networks primarily focuses on embedding secret information into social media text efficiently to achieve covert communication. The misuse of such techniques could pose significant potential threats to public cyberspace, such as the spread of malicious code, commands, or viruses. Existing social text steganalysis techniques mainly focus on the analysis of individual social media texts. However, the information content in a single text is very limited, leading to poor detection performance in practical applications. To address this challenge, this paper proposes a social text steganalysis method that combines large-scale language models with common-sense knowledge graphs (STLC-KG). This method first uses knowledge graphs to expand the knowledge contained in the text under investigation, enriching its linguistic expression, and then utilizes large-scale language models to extract the linguistic features of the social text. The results of tests conducted on three mainstream social media platforms demonstrate that the proposed method significantly improves the performance of social text steganalysis.

STLC-KG:A Social Text Steganalysis Method Combining Large-Scale Language Models and Common-Sense Knowledge Graphs

Local intrinsic dimension (LID) estimation methods have received a lot of attention in recent years thanks to the progress in deep neural networks and generative modeling. In opposition to old non-parametric methods, new methods use generative models to approximate diffused dataset density and scale the methods to high-dimensional datasets like images. In this paper, we investigate the recent state-of-the-art parametric LID estimation methods from the perspective of the Wiener process. We explore how these methods behave when their assumptions are not met. We give an extended mathematical description of those methods and their error as a function of the probability density of the data.

A Wiener Process Perspective on Local Intrinsic Dimension Estimation Methods

Large language models (LLMs) have made significant advancements in math problem solving, but their large size and high latency render them impractical for real-world applications in intelligent mathematics education. Recently, compact models have been developed to replace large LLMs in general natural language processing tasks. However, these models often struggle to acquire sufficient math-related knowledge from LLMs, leading to unsatisfactory performance in solving math word problems (MWPs). To develop a specialized compact model for solving math problems, we develop the knowledge distillation (KD) technique to distill mathematical semantic representations from BERT. Effective knowledge types and distillation strategies are explored through extensive experiments. Our KD algorithm employs multi-knowledge distillation to extract fundamental abstract knowledge from lower layers, as well as mathematics knowledge from higher layers, by leveraging bottleneck linear networks. Pre-training tasks, such as masked language modeling and part-of-speech tagging on MWP datasets, are also utilized to enhance generalization. Additionally, continual learning is employed to prevent catastrophic forgetting of acquired knowledge. Our findings indicate that our approach can reduce the size of a BERT model by 10% while retaining approximately 95% of its performance on MWP datasets, outperforming the mainstream BERT-based compact models. The efficacy of each component has been validated through ablation studies.

A Compact Model for Mathematics Problem Representations Distilled from BERT

With the widespread adoption of AI systems, many of the decisions once made by humans are now delegated to automated systems. Recent works in the literature demonstrate that these automated systems, when used in socially sensitive domains, may exhibit discriminatory behavior based on sensitive characteristics such as gender, sex, religion, or race. In light of this, various notions of fairness and methods to quantify discrimination have been proposed, also leading to the development of numerous approaches for constructing fair predictors. At the same time, imposing fairness constraints may decrease the utility of the decision-maker, highlighting a tension between fairness and utility. This tension is also recognized in legal frameworks, for instance in the disparate impact doctrine of Title VII of the Civil Rights Act of 1964 -- in which specific attention is given to considerations of \textit{business necessity} -- possibly allowing the usage of proxy variables associated with the sensitive attribute in case a high-enough utility cannot be achieved without them. In this work, we analyze the tension between fairness and accuracy from a causal lens for the first time. We introduce the notion of a path-specific excess loss (PSEL) that captures how much the predictor's loss increases when a causal fairness constraint is enforced. We then show that the total excess loss (TEL), defined as the difference between the loss of predictor fair along all causal pathways vs. an unconstrained predictor, can be decomposed into a sum of more local PSELs. At the same time, enforcing a causal constraint often reduces the disparity between demographic groups. Thus, we introduce a quantity that summarizes the fairness-utility trade-off, called the causal fairness/utility ratio, defined as the ratio of the reduction in discrimination vs. the excess in the loss from constraining a causal pathway. This quantity is particularly suitable for comparing the fairness-utility trade-off across different causal pathways. Finally, as our approach requires causally-constrained fair predictors, we introduce a new neural approach for causally-constrained fair learning. Our approach is evaluated across multiple real-world datasets, providing new insights into the tension between fairness and accuracy.

Fairness-Accuracy Trade-Offs: A Causal Perspective

Graph anomaly detection is crucial for identifying anomalous nodes within graphs and addressing applications like financial fraud detection and social spam detection. Recent spectral graph neural network methods advance graph anomaly detection by focusing on anomalies that notably affect the distribution of graph energy. Such spectrum-based methods rely on two steps: graph wavelet extraction and feature fusion. However, both steps are hand-designed, capturing incomprehensive anomaly information of wavelet-specific features and resulting in their inconsistent feature fusion. To address these problems, we propose a dynamic spectral graph anomaly detection framework DSGAD to adaptively capture comprehensive anomaly information and perform consistent feature fusion. DSGAD introduces dynamic wavelets, consisting of trainable wavelets to adaptively learn anomalous patterns and capture wavelet-specific features with comprehensive anomaly information. Furthermore, the consistent fusion of wavelet-specific features achieves dynamic fusion by combining wavelet-specific feature extraction with energy difference and channel convolution fusion using location correlation. Experimental results on four datasets substantiate the efficacy of our DSGAD method, surpassing state-of-the-art methods in both homogeneous and heterogeneous graphs. Our code will be published on GitHub.

Dynamic Spectral Graph Anomaly Detection

Messenger RNA (mRNA)-based vaccines are accelerating the discovery of new drugs and revolutionizing the pharmaceutical industry. However, selecting particular mRNA sequences for vaccines and therapeutics from extensive mRNA libraries is costly. Effective mRNA therapeutics require carefully designed sequences with optimized expression levels and stability. This paper proposes a novel contextual language model (LM)-based embedding method: mRNA2vec. In contrast to existing mRNA embedding approaches, our method is based on the self-supervised teacher-student learning framework of data2vec. We jointly use the 5' untranslated region (UTR) and coding sequence (CDS) region as the input sequences. We adapt our LM-based approach specifically to mRNA by 1) considering the importance of location on the mRNA sequence with probabilistic masking, 2) using Minimum Free Energy (MFE) prediction and Secondary Structure (SS) classification as additional pretext tasks. mRNA2vec demonstrates significant improvements in translation efficiency (TE) and expression level (EL) prediction tasks in UTR compared to SOTA methods such as UTR-LM. It also gives a competitive performance in mRNA stability and protein production level tasks in CDS such as CodonBERT.

NLSR: Neuron-Level Safety Realignment of Large Language Models Against Harmful Fine-Tuning

Downloads

Next from AAAI 2025

Faithful and Accurate Self-Attention Attribution for Message Passing Neural Networks via the Computation Tree Viewpoint