Akshaj Kumar Veldanda

Akshaj Kumar Veldanda, Fabian Grob, Shailja Thakur et al.

Large Language Models (LLMs) such as GPT-3.5, Bard, and Claude exhibit applicability across numerous tasks. One domain of interest is their use in algorithmic hiring, specifically in matching resumes with job categories. Yet, this introduces issues of bias on protected attributes like gender, race and maternity status. The seminal work of Bertrand & Mullainathan (2003) set the gold-standard for identifying hiring bias via field experiments where the response rate for identical resumes that differ only in protected attributes, e.g., racially suggestive names such as Emily or Lakisha, is compared. We replicate this experiment on state-of-art LLMs (GPT-3.5, Bard, Claude and Llama) to evaluate bias (or lack thereof) on gender, race, maternity status, pregnancy status, and political affiliation. We evaluate LLMs on two tasks: (1) matching resumes to job categories; and (2) summarizing resumes with employment relevant information. Overall, LLMs are robust across race and gender. They differ in their performance on pregnancy status and political affiliation. We use contrastive input decoding on open-source LLMs to uncover potential sources of bias.

Akshaj Kumar Veldanda, Ivan Brugere, Sanghamitra Dutta et al.

Fair machine learning methods seek to train models that balance model performance across demographic subgroups defined over sensitive attributes like race and gender. Although sensitive attributes are typically assumed to be known during training, they may not be available in practice due to privacy and other logistical concerns. Recent work has sought to train fair models without sensitive attributes on training data. However, these methods need extensive hyper-parameter tuning to achieve good results, and hence assume that sensitive attributes are known on validation data. However, this assumption too might not be practical. Here, we propose Antigone, a framework to train fair classifiers without access to sensitive attributes on either training or validation data. Instead, we generate pseudo sensitive attributes on the validation data by training a biased classifier and using the classifier's incorrectly (correctly) labeled examples as proxies for minority (majority) groups. Since fairness metrics like demographic parity, equal opportunity and subgroup accuracy can be estimated to within a proportionality constant even with noisy sensitive attribute information, we show theoretically and empirically that these proxy labels can be used to maximize fairness under average accuracy constraints. Key to our results is a principled approach to select the hyper-parameters of the biased classifier in a completely unsupervised fashion (meaning without access to ground truth sensitive attributes) that minimizes the gap between fairness estimated using noisy versus ground-truth sensitive labels.

Akshaj Kumar Veldanda, Ivan Brugere, Jiahao Chen et al.

The success of DNNs is driven by the counter-intuitive ability of over-parameterized networks to generalize, even when they perfectly fit the training data. In practice, test error often continues to decrease with increasing over-parameterization, referred to as double descent. This allows practitioners to instantiate large models without having to worry about over-fitting. Despite its benefits, however, prior work has shown that over-parameterization can exacerbate bias against minority subgroups. Several fairness-constrained DNN training methods have been proposed to address this concern. Here, we critically examine MinDiff, a fairness-constrained training procedure implemented within TensorFlow's Responsible AI Toolkit, that aims to achieve Equality of Opportunity. We show that although MinDiff improves fairness for under-parameterized models, it is likely to be ineffective in the over-parameterized regime. This is because an overfit model with zero training loss is trivially group-wise fair on training data, creating an "illusion of fairness," thus turning off the MinDiff optimization (this will apply to any disparity-based measures which care about errors or accuracy. It won't apply to demographic parity). Within specified fairness constraints, under-parameterized MinDiff models can even have lower error compared to their over-parameterized counterparts (despite baseline over-parameterized models having lower error). We further show that MinDiff optimization is very sensitive to choice of batch size in the under-parameterized regime. Thus, fair model training using MinDiff requires time-consuming hyper-parameter searches. Finally, we suggest using previously proposed regularization techniques, viz. L2, early stopping and flooding in conjunction with MinDiff to train fair over-parameterized models.

Akshaj Kumar Veldanda, Shi-Xiong Zhang, Anirban Das et al.

Large language models (LLMs) have revolutionized various domains, yet their utility comes with significant challenges related to outdated or problematic knowledge embedded during pretraining. This paper addresses the challenge of modifying LLMs to unlearn problematic and outdated information while efficiently integrating new knowledge without retraining from scratch. Here, we propose LLM Surgery, a framework to efficiently modify LLM behaviour by optimizing a three component objective function that: (1) Performs reverse gradient on unlearning dataset (problematic and outdated information), (2) Performs gradient descent on the update dataset (new and updated information), and (3) Minimizes the KL divergence on the retain dataset (small subset of unchanged text), ensuring alignment between pretrained and modified model outputs. Due to the lack of publicly available datasets specifically tailored for our novel task, we compiled a new dataset and an evaluation benchmark. Using Llama2-7B, we demonstrate that LLM Surgery can achieve significant forgetting on the unlearn set, a 20\% increase in accuracy on the update set, and maintain performance on the retain set.

Swagnik Roychoudhury, Akshaj Kumar Veldanda

Spam filters are a crucial component of modern email systems, as they help to protect users from unwanted and potentially harmful emails. However, the effectiveness of these filters is dependent on the quality of the machine learning models that power them. In this paper, we design backdoor attacks in the domain of spam filtering. By demonstrating the potential vulnerabilities in the machine learning model supply chain, we highlight the need for careful consideration and evaluation of the models used in spam filters. Our results show that the backdoor attacks can be effectively used to identify vulnerabilities in spam filters and suggest the need for ongoing monitoring and improvement in this area.

Hao Fu, Akshaj Kumar Veldanda, Prashanth Krishnamurthy et al.

This paper proposes a new defense against neural network backdooring attacks that are maliciously trained to mispredict in the presence of attacker-chosen triggers. Our defense is based on the intuition that the feature extraction layers of a backdoored network embed new features to detect the presence of a trigger and the subsequent classification layers learn to mispredict when triggers are detected. Therefore, to detect backdoors, the proposed defense uses two synergistic anomaly detectors trained on clean validation data: the first is a novelty detector that checks for anomalous features, while the second detects anomalous mappings from features to outputs by comparing with a separate classifier trained on validation data. The approach is evaluated on a wide range of backdoored networks (with multiple variations of triggers) that successfully evade state-of-the-art defenses. Additionally, we evaluate the robustness of our approach on imperceptible perturbations, scalability on large-scale datasets, and effectiveness under domain shift. This paper also shows that the defense can be further improved using data augmentation.

Akshaj Kumar Veldanda, Kang Liu, Benjamin Tan et al.

This paper proposes a novel two-stage defense (NNoculation) against backdoored neural networks (BadNets) that, repairs a BadNet both pre-deployment and online in response to backdoored test inputs encountered in the field. In the pre-deployment stage, NNoculation retrains the BadNet with random perturbations of clean validation inputs to partially reduce the adversarial impact of a backdoor. Post-deployment, NNoculation detects and quarantines backdoored test inputs by recording disagreements between the original and pre-deployment patched networks. A CycleGAN is then trained to learn transformations between clean validation and quarantined inputs; i.e., it learns to add triggers to clean validation images. Backdoored validation images along with their correct labels are used to further retrain the pre-deployment patched network, yielding our final defense. Empirical evaluation on a comprehensive suite of backdoor attacks show that NNoculation outperforms all state-of-the-art defenses that make restrictive assumptions and only work on specific backdoor attacks, or fail on adaptive attacks. In contrast, NNoculation makes minimal assumptions and provides an effective defense, even under settings where existing defenses are ineffective due to attackers circumventing their restrictive assumptions.