$\textit{Who Speaks Matters}$: Analysing the Influence of the Speaker's Ethnicity on Hate Classification

Large Language Models (LLMs) offer a lucrative promise for scalable content moderation, including hate speech detection. However, they are also known to be brittle and biased against marginalised communities and dialects. This requires their applications to high-stakes tasks like hate speech detection to be critically scrutinized. In this work, we investigate the robustness of hate speech classification using LLMs, particularly when explicit and implicit markers of the speaker's ethnicity are injected into the input. For the explicit markers, we inject a phrase that mentions the speaker's identity. For the implicit markers, we inject dialectal features. By analysing how frequently model outputs flip in the presence of these markers, we reveal varying degrees of brittleness across 4 popular LLMs and 5 ethnicities. We find that the presence of implicit dialect markers in inputs causes model outputs to flip more than the presence of explicit markers. Further, the percentage of flips varies across ethnicities. Finally, we find that larger models are more robust. Our findings indicate the need for exercising caution in deploying LLMs for high-stakes tasks like hate speech detection.

Mysterious Projections: Multimodal LLMs Gain Domain-Specific Visual Capabilities Without Richer Cross-Modal Projections

Multimodal large language models (MLLMs) like LLaVA and GPT-4(V) enable general-purpose conversations about images with the language modality. As off-the-shelf MLLMs may have limited capabilities on images from domains like dermatology and agriculture, they must be fine-tuned to unlock domain-specific applications. The prevalent architecture of current open-source MLLMs comprises two major modules: an image-language (cross-modal) projection network and a large language model. It is desirable to understand the roles of these two modules in modeling domain-specific visual attributes to inform the design of future models and streamline the interpretability efforts on the current models. To this end, via experiments on 4 datasets and under 2 fine-tuning settings, we find that as the MLLM is fine-tuned, it indeed gains domain-specific visual capabilities, but the updates do not lead to the projection extracting relevant domain-specific visual attributes. Our results indicate that the domain-specific visual attributes are modeled by the LLM, even when only the projection is fine-tuned. Through this study, we offer a potential reinterpretation of the role of cross-modal projections in MLLM architectures. Projection webpage: https://claws-lab.github.io/projection-in-MLLMs/

A Survey on Explainability of Graph Neural Networks

Graph neural networks (GNNs) are powerful graph-based deep-learning models that have gained significant attention and demonstrated remarkable performance in various domains, including natural language processing, drug discovery, and recommendation systems. However, combining feature information and combinatorial graph structures has led to complex non-linear GNN models. Consequently, this has increased the challenges of understanding the workings of GNNs and the underlying reasons behind their predictions. To address this, numerous explainability methods have been proposed to shed light on the inner mechanism of the GNNs. Explainable GNNs improve their security and enhance trust in their recommendations. This survey aims to provide a comprehensive overview of the existing explainability techniques for GNNs. We create a novel taxonomy and hierarchy to categorize these methods based on their objective and methodology. We also discuss the strengths, limitations, and application scenarios of each category. Furthermore, we highlight the key evaluation metrics and datasets commonly used to assess the explainability of GNNs. This survey aims to assist researchers and practitioners in understanding the existing landscape of explainability methods, identifying gaps, and fostering further advancements in interpretable graph-based machine learning.

NoisyTwins: Class-Consistent and Diverse Image Generation through StyleGANs

StyleGANs are at the forefront of controllable image generation as they produce a latent space that is semantically disentangled, making it suitable for image editing and manipulation. However, the performance of StyleGANs severely degrades when trained via class-conditioning on large-scale long-tailed datasets. We find that one reason for degradation is the collapse of latents for each class in the $\mathcal{W}$ latent space. With NoisyTwins, we first introduce an effective and inexpensive augmentation strategy for class embeddings, which then decorrelates the latents based on self-supervision in the $\mathcal{W}$ space. This decorrelation mitigates collapse, ensuring that our method preserves intra-class diversity with class-consistency in image generation. We show the effectiveness of our approach on large-scale real-world long-tailed datasets of ImageNet-LT and iNaturalist 2019, where our method outperforms other methods by $\sim 19\%$ on FID, establishing a new state-of-the-art.

A Survey of Graph Neural Networks for Social Recommender Systems

Social recommender systems (SocialRS) simultaneously leverage user-to-item interactions as well as user-to-user social relations for the task of generating item recommendations to users. Additionally exploiting social relations is clearly effective in understanding users' tastes due to the effects of homophily and social influence. For this reason, SocialRS has increasingly attracted attention. In particular, with the advance of Graph Neural Networks (GNN), many GNN-based SocialRS methods have been developed recently. Therefore, we conduct a comprehensive and systematic review of the literature on GNN-based SocialRS. In this survey, we first identify 80 papers on GNN-based SocialRS after annotating 2151 papers by following the PRISMA framework (Preferred Reporting Items for Systematic Reviews and Meta-Analysis). Then, we comprehensively review them in terms of their inputs and architectures to propose a novel taxonomy: (1) input taxonomy includes 5 groups of input type notations and 7 groups of input representation notations; (2) architecture taxonomy includes 8 groups of GNN encoder, 2 groups of decoder, and 12 groups of loss function notations. We classify the GNN-based SocialRS methods into several categories as per the taxonomy and describe their details. Furthermore, we summarize the benchmark datasets and metrics widely used to evaluate the GNN-based SocialRS methods. Finally, we conclude this survey by presenting some future research directions.

Signed Link Representation in Continuous-Time Dynamic Signed Networks

Signed networks allow us to model bi-faceted relationships and interactions, such as friend/enemy, support/oppose, etc. These interactions are often temporal in real datasets, where nodes and edges appear over time. Learning the dynamics of signed networks is thus crucial to effectively predict the sign and strength of future links. Existing works model either signed networks or dynamic networks but not both together. In this work, we study dynamic signed networks where links are both signed and evolving with time. Our model learns a Signed link's Evolution using Memory modules and Balanced Aggregation (hence, the name SEMBA). Each node maintains two separate memory encodings for positive and negative interactions. On the arrival of a new edge, each interacting node aggregates this signed information with its memories while exploiting balance theory. Node embeddings are generated using updated memories, which are then used to train for multiple downstream tasks, including link sign prediction and link weight prediction. Our results show that SEMBA outperforms all the baselines on the task of sign prediction by achieving up to an 8% increase in the AUC and up to a 50% reduction in FPR. Results on the task of predicting signed weights show that SEMBA reduces the mean squared error by 9% while achieving up to 69% reduction in the KL-divergence on the distribution of predicted signed weights.

Task and Model Agnostic Adversarial Attack on Graph Neural Networks

Graph neural networks (GNNs) have witnessed significant adoption in the industry owing to impressive performance on various predictive tasks. Performance alone, however, is not enough. Any widely deployed machine learning algorithm must be robust to adversarial attacks. In this work, we investigate this aspect for GNNs, identify vulnerabilities, and link them to graph properties that may potentially lead to the development of more secure and robust GNNs. Specifically, we formulate the problem of task and model agnostic evasion attacks where adversaries modify the test graph to affect the performance of any unknown downstream task. The proposed algorithm, GRAND ($Gr$aph $A$ttack via $N$eighborhood $D$istortion) shows that distortion of node neighborhoods is effective in drastically compromising prediction performance. Although neighborhood distortion is an NP-hard problem, GRAND designs an effective heuristic through a novel combination of Graph Isomorphism Network with deep $Q$-learning. Extensive experiments on real datasets show that, on average, GRAND is up to $50\%$ more effective than state of the art techniques, while being more than $100$ times faster.

Hiding Data in Images Using Cryptography and Deep Neural Network

Steganography is an art of obscuring data inside another quotidian file of similar or varying types. Hiding data has always been of significant importance to digital forensics. Previously, steganography has been combined with cryptography and neural networks separately. Whereas, this research combines steganography, cryptography with the neural networks all together to hide an image inside another container image of the larger or same size. Although the cryptographic technique used is quite simple, but is effective when convoluted with deep neural nets. Other steganography techniques involve hiding data efficiently, but in a uniform pattern which makes it less secure. This method targets both the challenges and make data hiding secure and non-uniform.