United States

We report a new neural backdoor attack, named Hibernated Backdoor, which is stealthy, aggressive and devastating. The backdoor is planted in a hibernated mode to avoid being detected. Once deployed and fine-tuned on end-devices, the hibernated backdoor turns into the active state that can be exploited by the attacker. To the best of our knowledge, this is the first hibernated neural backdoor attack. It is achieved by maximizing the mutual information (MI) between the gradients of regular and malicious data on the model. We introduce a practical algorithm to achieve MI maximization to effectively plant the hibernated backdoor. To evade adaptive defenses, we further develop a targeted hibernated backdoor, which can only be activated by specific data samples and thus achieves a higher degree of stealthiness. We show the hibernated backdoor is robust and cannot be removed by existing backdoor removal schemes. It has been fully tested on four datasets with two neural network architectures, compared to five existing backdoor attacks, and evaluated using seven backdoor detection schemes. The experiments demonstrate the effectiveness of the hibernated backdoor attack under various settings.

AAAI 2022

Hibernated Backdoor: A Mutual Information Empowered Backdoor Attack to Deep Neural Networks

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community."<br>

*Message from the General Chair, Dirk Hovy*





<br>

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

technical paper

The purpose of the AAAI conference is to promote research in artificial intelligence (AI) and scientific exchange among AI researchers, practitioners, scientists, and engineers in affiliated disciplines. 

Recent studies in the hate and counter hate domain have provided the grounds for investigating how to detect this pervasive content in social media. These studies mostly work with synthetic replies to hateful content written by annotators on demand rather than replies written by real users. We argue that working with naturally occurring replies to hateful content is key to study the problem. Building on this motivation, we create a corpus of 5,652 hateful tweets and replies. We analyze their fine-grained relationships by indicating whether the reply (a) is hate or counter hate speech, (b) provides a justification, (c) attacks the author of the tweet, and (d) adds additional hate. We also present linguistic insights into the language people use depending on these fine-grained relationships. Experimental results show improvements (a) taking into account the hateful tweet in addition to the reply and (b) pretraining with related tasks.

Pinpointing Fine-Grained Relationships between Hateful Tweets and Replies

We explore the utility of information contained within a dropout based Bayesian neural network (BNN) for the task of detecting out of distribution (OOD) data. We first show how previous attempts to leverage the randomized embeddings induced by the intermediate layers of a dropout BNN can fail due to the distance metric used. We introduce an alternative approach to measuring embedding uncertainty, and demonstrate how incorporating embedding uncertainty improves OOD data identification across three tasks: image classification, language classification, and malware detection.

Out of Distribution Data Detection Using Dropout Bayesian Neural Networks

Detecting rumors on social media has become particular important due to the rapid dissemination and adverse impacts on our lives. Though a set of rumor detection models have exploited the message propagation structural or temporal information, they seldom model them altogether to enjoy the best of both worlds. Moreover, the dynamics of knowledge information associated with the comments are not involved, either. To this end, we propose a novel Dual-Dynamic Graph Convolutional Networks, termed as DDGCN, which can model the dynamics of messages in propagation as well as the dynamics of the background knowledge from Knowledge graphs in one unified framework. Specifically, two Graph Convolutional Networks are adopted to capture the above two types of structure information at different time stages, which are then combined with a temporal fusing unit. This allows for learning the dynamic event representations in a more fine-grained manner, and incrementally aggregating them to capture the cascading effect for better rumor detection. Extensive experiments on two public real-world datasets demonstrate that our proposal yields significant improvements compared to strong baselines and can detect rumors at early stages.

DDGCN: Dual Dynamic Graph Convolutional Networks for Rumor Detection on Social Media

Pre-training language models (LMs) on large-scale unlabeled text data makes the model much easier to achieve exceptional downstream performance than their counterparts directly trained on the downstream tasks. 
 In this work, we study what specific traits in the pre-training data, other than the semantics, make a pre-trained LM superior to their counterparts trained from scratch on downstream tasks.
 We propose to use artificially constructed datasets as the pre-training data to exclude the effect of semantics, and further control what characteristics the pre-training corpora have.
 By fine-tuning the pre-trained models on GLUE benchmark, we can learn how beneficial it is to transfer the knowledge from the model trained on the dataset possessing that specific trait.
 We define and discuss three different characteristics in the artificial dataset: 1) matching the token's uni-gram or bi-gram distribution between pre-training and downstream fine-tuning, 2) the presence of the explicit dependencies among the tokens in a sequence, 3) the length of the implicit dependencies among the tokens in a sequence. 
 Our experiments show that the explicit dependencies in the sequences of the pre-training data are critical to the downstream performance.
 Our results also reveal that models achieve better downstream performance when pre-trained on a dataset with a longer range of implicit dependencies.
 Based on our analysis, we find that models pre-trained with artificial datasets are prone to learn spurious correlation in downstream tasks.
 Our work reveals that even if the LMs are not pre-trained on natural language, they still gain transferability on certain human language downstream tasks once the LMs learn to model the token dependencies in the sequences. 
 This result helps us understand the exceptional transferability of pre-trained LMs.

On the Transferability of Pre-Trained Language Models: A Study from Artificial Datasets

Neural sequence models trained with maximum likelihood estimation have led to breakthroughs in many tasks, where success is defined by the gap between training and test performance. However, their ability to achieve stronger forms of generalization remains unclear. We consider the problem of symbolic mathematical integration, as it requires generalizing systematically beyond the training set. We develop a methodology for evaluating generalization that takes advantage of the problem domain's structure and access to a verifier. Despite promising in-distribution performance of sequence-to-sequence models in this domain, we demonstrate challenges in achieving robustness, compositionality, and out-of-distribution generalization, through both carefully constructed manual test suites and a genetic algorithm that automatically finds large collections of failures in a controllable manner. Our investigation highlights the difficulty of generalizing well with the predominant modeling and learning approach, and the importance of evaluating beyond the test set, across different aspects of generalization.

Symbolic Brittleness in Sequence Models: On Systematic Generalization in Symbolic Mathematics

We propose a dual-domain generative model to estimate a texture map from a single image for colorizing a 3D human model. When estimating a texture map, a single image is insufficient as it reveals only one facet of a 3D object. To provide sufficient information for estimating a complete texture map, the proposed model simultaneously generates multi-view hallucinations in the image domain and an estimated texture map in the texture domain. During the generating process, each domain generator exchanges features to the other by a flow-based local attention mechanism. In this manner, the proposed model can estimate a texture map utilizing abundant multi-view image features from which multiview hallucinations are generated. As a result, the estimated texture map contains consistent colors and patterns over the entire region. Experiments show the superiority of our model for estimating a directly render-able texture map, which is applicable to 3D animation rendering. Furthermore, our model also improves an overall generation quality in the image domain for pose and viewpoint transfer tasks.

Texture Generation Using Dual-Domain Feature Flow with Multi-View Hallucinations

We consider language modelling (LM) as a multi-label structured prediction task by re-framing training from solely predicting a single ground-truth word to ranking a set of words which could continue a given context. To avoid annotating top-k ranks, we generate them using pre-trained LMs: GPT-2, BERT, and Born-Again models. This leads to a rank-based form of knowledge distillation (KD). We also develop a method using N-grams to create a non-probabilistic teacher which generates the ranks without the need of a pre-trained LM.
 
 We confirm the hypotheses: that we can treat LMing as a ranking task and that we can do so without the use of a pre-trained LM.
 We show that rank-based KD generally gives a modest improvement to perplexity (PPL) -- though often with statistical significance -- when compared to Kullbackâ€“Leibler-based KD. Surprisingly, given the naivety of the method, the N-grams act as competitive teachers and achieve similar performance as using either BERT or a Born-Again model teachers. Unsurprisingly, GPT-2 always acts as the best teacher. 
 Using it and a Transformer-XL student on Wiki-02, rank-based KD reduces a cross-entropy baseline from 65.27 to 55.94 and against a KL-based KD of 56.70.

Language Modelling via Learning to Rank

Existential rule languages are a family of ontology languages that have been widely used in ontology-mediated query answering (OMQA). However, for most of them, the expressive power of representing domain knowledge for OMQA, known as the program expressive power, is not well-understood yet. In this paper, we establish a number of novel characterizations for the program expressive power of several important existential rule languages, including tuple-generating dependencies (TGDs), linear TGDs, as well as disjunctive TGDs. The characterizations employ natural model-theoretic properties, and automata-theoretic properties sometimes, which thus provide powerful tools for identifying the definability of domain knowledge for OMQA in these languages.

Characterizing the Program Expressive Power of Existential Rule Languages

An important problem across multiple disciplines is to infer and understand meaningful latent variables. One strategy commonly used is to model the measured variables in terms of the latent variables under suitable assumptions on the connectivity from the latents to the measured (known as measurement model). Furthermore, it might be even more interesting to discover the causal relations among the latent variables (known as structural model). Recently, some methods have been proposed to estimate the structural model by assuming that the noise terms in the measured and latent variables are non-Gaussian. However, they are not suitable when some of the noise terms become Gaussian. To bridge this gap, we investigate the problem of identification of the structural model with arbitrary noise distributions. We provide necessary and sufficient condition under which the structural model is identifiable: it is identifiable iff for each pair of adjacent latent variables Lx, Ly, (1) at least one of Lx and Ly has non-Gaussian noise, or (2) at least one of them has a non-Gaussian ancestor and is not d-separated from the non-Gaussian component of this ancestor by the common causes of Lx and Ly. This identifiability result relaxes the non-Gaussianity requirements to only a (hopefully small) subset of variables, and accordingly elegantly extends the application scope of the structural model. Based on the above identifiability result, we further propose a practical algorithm to learn the structural model. We verify the correctness of the identifiability result and the effectiveness of the proposed method through empirical studies.

Identification of Linear Latent Variable Model with Arbitrary Distribution

Commonsense question answering (CQA) aims to test if models can answer questions regarding commonsense knowledge that everyone knows. Prior works that incorporate external knowledge bases have shown promising results, but knowledge bases are expensive to construct and are often limited to a fixed set of relations. In this paper, we instead focus on better utilizing the implicit knowledge stored in pre-trained language models. While researchers have found that the knowledge embedded in pre-trained language models can be extracted by having them fill in the blanks of carefully designed prompts for relation extraction and text classification, it remains unclear if we can adopt this paradigm in CQA where the inputs and outputs take much more flexible forms. To this end, we investigate four translation methods that can translate natural questions into cloze-style sentences to better solicit commonsense knowledge from language models, including a syntactic-based model, an unsupervised neural model, and two supervised neural models. In addition, to combine the different translation methods, we propose to encourage consistency among model predictions on different translated questions with unlabeled data. We demonstrate the effectiveness of our methods on three CQA datasets in zero-shot settings. We show that our methods are complementary to a knowledge base improved model, and combining them can lead to state-of-the-art zero-shot performance. Analyses also reveal distinct characteristics of the different cloze translation methods and provide insights on why combining them can lead to great improvements. Code/dataset is available at https://github.com/PlusLabNLP/zero_shot_cqa.

Downloads

Next from AAAI 2022

Pinpointing Fine-Grained Relationships between Hateful Tweets and Replies

Similar lecture

Forget the Unneeded: Backdooring Large Language Models via Contrastive-enhanced Machine Unlearning

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from AAAI 2022

Pinpointing Fine-Grained Relationships between Hateful Tweets and Replies

Similar lecture

Forget the Unneeded: Backdooring Large Language Models via Contrastive-enhanced Machine Unlearning

Downloads