Abstract

Model Merging (MM) has emerged as a promising approach to combine multiple fine-tuned models into a single model that maintains performance across tasks. However, its impact on (transferable) adversarial examples remains unexplored. Specifically, we investigate the affect of MM on the transferability of adversarial examples---a black-box, adversarial attack where adversarial examples generated for surrogate model successfully mislead target model. Through experiments on image classification models, we discover that, similar to recent findings of MM on backdoor attacks, where MM tends to sanitize the backdoor present in individual fine-tuned models, MM does not boost adversarial transferability. Our analysis reveals MM has no affect, and under certain conditions, decreased vulnerability to transferable adversarial examples. Consequently, our finding also support the notion of MM giving a "free lunch'' of adversarial robustness and provides important insights for designing more secure systems that employ MM.

Abstract

According to the cooperative principle (Grice 1975), utterances in a conversation adhere to the four maxims of conversational implicature – quantity, relation, quality and manner. An implicature arises when a maxim is flouted or contradicted, and the speaker is forced to infer a meaning. Consider the example: Speaker A: का िदलीप को परीकाएं आसान लग रही है? Kya Dilip ko parikshayein aasan lag rahi hain? (Does Dilip find the exam easy?) Speaker B: आज-कल वो सो नही पा रहा | [Aaj-kal wo so nahi pa raha.] (He hasn’t been able to sleep lately.) Speaker B flouts the maxim of relevance by not answering the question with a yes or no. Instead, his reply that Dilip has trouble sleeping implies that he doesn’t find the exams to be easy. While computational approaches have made some progress in English implicature benchmarks - GRICE (Zheng, 2021), IMPRESS (Jeretic 2020) and Conversational Implicature benchmark (CIB) (George and Mamidi, 2020), there are no benchmarks for Hindi. Hence, we propose a Hindi implicature benchmark- Indirect Questions Hi Implicature Benchmark - Toledo-Ronen (2020) found that translation- based methods lose nuances in argument evaluation. XCOPA benchmark (Ponti 2020) & Huang (2024) demonstrated that translated data hurts model performance for structurally divergent non- English languages. Hindi is a free word order, morphologically rich language. It is structurally different from English hence, translating English implicature benchmarks may not be unsuitable. Wang (2024) present indirect answers to yes-no questions in dialogues benchmark. Similar to their approach, we plan to collect Hindi interviews and extract indirect questions along with their answers involving implicature. We will annotate 5000 such questions. We will maintain a constraint that previous dialogues will be added to the question as a background to simplify the inference over a single dialogue turn. Evaluation will rely on ROUGE scores to measure correctness, as we leave exploration of suitable metrics for future work. For model evaluation, we plan to leverage a) English LLMs (meta-llama/Llama-3.3-70B (Grattafiori 2024) and the current state-of-the-art reported by Sravanthi (2024), FlanT5-XXL (Chung 2022) finetuned on Hindi translated GRICE (Zheng 2021), IMPRESS (Jeretic 2020) and CIB (George and Mamidi 2020) and b) Hindi LLMs (Llama-3-Nanda-10B-Chat (Choudhury 2024) and Airavata (Gala 2024)).

Abstract

In recent years, widespread internet adoption and the growth in user base of various social media platforms have led to an increase in the prolifera tion of extreme speech online. While traditional language models have demon strated proficiency in distinguishing between neutral text and non-neutral text (i.e. extreme speech), categorizing the diverse types of extreme speech presents significant challenges [1][2]. The task of extreme speech classification is par ticularly nuanced, as it requires a deep understanding of socio-cultural contexts to accurately interpret the intent of the language used by the speaker. Even hu man annotators often disagree on the appropriate classification of such content, emphasizing the complex and subjective nature of this task [3]. The use of hu man moderators also presents a scaling issue, necessitating the need for auto mated systems for extreme speech classification. The recent launch of ChatGPT has drawn global attention to the potential applications of Large Language Models (LLMs) across a diverse variety of tasks. Trained on vast and diverse corpora, and demonstrating the ability to effectively capture and encode contex tual information, LLMs emerge as highly promising tools for tackling this spe cific task of extreme speech classification. In this paper, we leverage the Indian subset of the extreme speech dataset from [3] to develop an effective classifica tion framework using LLMs. We evaluate open-source Llama models against closed-source OpenAI models, finding that while pre-trained LLMs show mod erate efficacy, fine-tuning with domain-specific data significantly enhances per formance, highlighting their adaptability to linguistic and contextual nuances. Although GPT-based models outperform Llama models in zero-shot settings, the performance gap disappears after fine-tuning.

Abstract

Modeling process, operating condition, reliability, and technological variation in transistor level design introduce significant variance in device and performance parameters. This paper presents a machine-learning-based architectural approach that enhances model performance, as a step towards developing Foundation Models tailored for circuit data, accurately capturing technology and process-induced variations in leakage power, voltage, and current. The architecture incorporates the impact of varying operating conditions, including temperatures from -55°C to 125°C and supply voltage fluctuations of ±10% on 16nm HP FinFET, and 16nm, 22nm, 32nm, and 45nm HPMGK CMOS technology nodes. By enhancing the performance of baseline machine-learning models using a chained architecture to cater to the high variance in target, our C-Arch serves as a versatile framework for circuit modeling when the data spread is high. Experimental results on current, voltage, and leakage power estimation, demonstrate average improvements of up to 93.76%, 96.17%, and 88.82% in Mean Absolute Percentage Error compared to baseline models, underscoring the computational savings and viability of Foundation Models in circuit design.

Abstract

We consider a reconfigurable intelligent surface (RIS)-assisted multi-user communication system. For such a system, we aim to optimally select the RIS phase shifts and precoding vectors for maximizing the effective rank of the weighted channel covariance matrix which essentially improves the channel capacity. For a low-complexity transmitter design, we employ maximum ratio transmission (MRT) and minimum-mean square error (MMSE) precoding schemes along with water-filling algorithm-based power allocation. Further, we show that MRT and MMSE exhibit equivalent performance and become optimal when the channel effective rank is maximized by optimally configuring the RIS consisting of a large number of elements

Abstract

Graph Neural Networks (GNNs) are increasingly being used for a variety of ML applications on graph data. Because graph data does not follow the independently and identically distributed i.i.d. assumption, adversarial manipulations or incorrect data can propagate to other data points through message passing, which deteriorates the model's performance. To allow model developers to remove the adverse effects of manipulated entities from a trained GNN, we study the recently formulated problem of Corrective Unlearning. We find that current graph unlearning methods fail to unlearn the effect of manipulations even when the whole manipulated set is known. We introduce a new graph unlearning method,Cognac, which can unlearn the effect of the manipulation set even when only % of it is identified. It recovers most of the performance of a strong oracle with fully corrected training data, even beating retraining from scratch without the deletion set, and is 8x more efficient while also scaling to large datasets. We hope our work assists GNN developers in mitigating harmful effects caused by issues in real-world data, post-training.

Abstract

Non-Structural Elements (NSEs) in buildings are expected to remain operational after earthquakes. But, numerical models of Structural Elements (SEs) (buildings) seldom include NSEs, thereby ignoring dynamic interaction between SE-NSEs, resulting in unrealistic earthquake demands. This study aims to quantify these earthquake demands, especially in acceleration-sensitive NSEs, by using Floor Response Spectrum (FRS) approach, and employing Decoupled & Coupled Models during Nonlinear Response History Analysis. Results are examined for effectiveness and accuracy in predicting acceleration demands. The findings emphasize importance of incorporating coupled SE-NSE behavior, for improved seismic design and behavior of buildings and NSEs during earthquake events.

Abstract

The Facility Location Problem (FLP) is a well-studied optimization problem with applications in many real-world scenarios. Past literature has explored the solutions from different perspectives to tackle FLPs. These include investigating FLPs under objective functions such as utilitarian, egalitarian, Nash welfare, etc. Also, there is no treatment for asymmetric welfare functions around the facility. We propose a unified framework, FLIGHT, to accommodate a broad class of welfare notions. The framework undergoes rigorous theoretical analysis, and we prove some structural properties of the solution to FLP. Additionally, we provide approximation bounds, which provide insight into an interesting fact: as the number of agents arbitrarily increases, the choice of welfare notion is irrelevant. Furthermore, the paper also includes results around concentration bounds under certain distributional assumptions over the preferred locations of agents.

Abstract

This paper considers the problem of fair probabilistic binary classification with binary protected groups. The classifier assigns scores, and a practitioner predicts labels using a certain cut-off threshold based on the desired trade-off between false positives vs. false negatives. It derives these thresholds from the ROC of the classifier. The resultant classifier may be unfair to one of the two protected groups in the dataset. It is desirable that no matter what threshold the practitioner uses, the classifier should be fair to both the protected groups; that is, the ℒₚ norm between FPRs and TPRs of both the protected groups should be at most ε. We call such fairness on ROCs of both the protected attributes εₚ-Equalized ROC. Given a classifier not satisfying ε₁-Equalized ROC, we aim to design a post-processing method to transform the given (potentially unfair) classifier's output (score) to a suitable randomized yet fair classifier. That is, the resultant classifier must satisfy ε₁-Equalized ROC. First, we introduce a threshold query model on the ROC curves for each protected group. The resulting classifier is bound to face a reduction in AUC. With the proposed query model, we provide a rigorous theoretical analysis of the minimal AUC loss to achieve ε₁-Equalized ROC. To achieve this, we design a linear time algorithm, namely FROC, to transform a given classifier's output to a probabilistic classifier that satisfies ε₁-Equalized ROC. We prove that under certain theoretical conditions, FROC achieves the theoretical optimal guarantees. We also study the performance of our FROC on multiple real-world datasets with many trained classifiers.

Abstract

Re-identification (ReID) is a critical challenge in computer vision, predominantly studied in the context of pedestrians and vehicles. However, robust object-instance ReID, which has significant im- plications for tasks such as autonomous exploration, long-term perception, and scene understanding, remains underexplored. In this work, we address this gap by proposing a novel dual-path object-instance re-identification transformer architecture that inte- grates multimodal RGB and depth information. By leveraging depth data, we demonstrate improvements in ReID across scenes that are cluttered or have varying illumination conditions. Additionally, we develop a ReID-based localization framework that enables accurate camera localization and pose identification across different view- points. We validate our methods using two custom-built RGB-D datasets, as well as multiple sequences from the open-source TUM RGB-D datasets. Our approach demonstrates significant improve- ments in both object instance ReID (mAP of 75.18) and localization accuracy (success rate of 83% on TUM-RGBD), highlighting the essential role of object ReID in advancing robotic perception. Our models, frameworks, and datasets have been made publicly avail- able.