Blog posts

2024

Mitigating Biases with Adversarial Learning in Computer Vision

August 11, 2024

In the era of machine learning and artificial intelligence, biases in models have emerged as a significant concern. These biases can lead to unfair treatment of individuals based on race, gender, or other characteristics, which is particularly problematic in applications like facial recognition, hiring algorithms, and criminal justice. Adversarial learning is a promising approach to mitigate these biases while maintaining model performance. This blog will provide an overview of biases in machine learning, demonstrate training a computer vision network, evaluate its biases, and illustrate how adversarial learning can mitigate these biases without compromising performance.

Understanding Data Poisoning in the Realm of Adversarial Attacks

January 01, 2024

Explore the realm of adversarial attacks on machine learning models in this insightful blog. Beginning with an introduction to data poisoning, we look into practical demonstrations using linear regression and PyTorch-based neural networks. The blog also discusses essential strategies for detection and mitigation, emphasizing robust model training, ensemble methods, data sanitization, differential privacy, and outlier rejection. As a comprehensive guide, it underscores the importance of ongoing research and vigilance in fortifying machine learning systems against evolving adversarial threats.

2023

Adversarial Attacks with Carlini & Wagner Approach

December 28, 2023

In this exploration of adversarial attacks, we unveil the Carlini & Wagner (C&W) approach, a potent method for crafting imperceptible but deceptive inputs. We introduce the concept of adversarial attacks, emphasizing their impact on the vulnerability of machine learning models. Our journey begins with setting up and training a neural network, paving the way for implementing the C&W attack. The effectiveness of the attack is showcased through visualizations, demonstrating the subtle yet powerful manipulations it imposes on model predictions. The blog concludes by stressing the critical need to comprehend and address adversarial threats in the pursuit of robust and secure AI systems.

Unveiling the Power of Projected Gradient Descent in Adversarial Attacks

December 26, 2023

Unveiling the intricacies of adversarial attacks in machine learning, this article explores Projected Gradient Descent (PGD) as a pivotal optimization algorithm. PGD, building on the foundation of gradient descent, excels in crafting robust adversarial examples, offering transferability and stability. Despite computational costs and hyperparameter considerations, PGD stands as a powerful tool for evaluating model vulnerabilities. The article discusses its advantages, such as generating universal perturbations, and its limitations, including sensitivity to hyperparameters. Navigating the delicate balance between accuracy and resilience, understanding PGD emerges as crucial for fortifying machine learning models against evolving adversarial landscapes.

A Deep Dive into the Fast Gradient Sign Method

December 23, 2023

Dive into the realm of adversarial machine learning with this blog post. Unraveling the Fast Gradient Sign Method (FGSM), we explore the vulnerability of neural networks to subtle perturbations. Illustrated by the iconic Panda Experiment from Ian Goodfellow's research, our blog elucidates how FGSM manipulates inputs to mislead models without human detection. The mathematical intricacies behind FGSM are unveiled, showcasing the power and potential risks of adversarial attacks in artificial intelligence. With a concise PyTorch code snippet and an invitation to a Google Colab notebook, readers can witness FGSM in action, perturbing CIFAR-10 dataset inputs to provoke incorrect predictions

Accelerating Machine Learning with Distributed XGBoost: A Tutorial

September 01, 2023

XGBoost (which stands for eXtreme Gradient Boosting) is a popular machine learning algorithm that is widely used for building high-performance models. In this tutorial, we will show you how to use XGBoost in a distributed environment to train models faster and more efficiently than ever before. By distributing the workload of training a model across multiple machines, we can process larger datasets and train more complex models in a fraction of the time it would take on a single machine. We will walk you through the process of preparing your data for training and running XGBoost in distributed mode. Whether you are working with large datasets or tackling complex problems, using XGBoost in a distributed environment can help you achieve better results in less time.

2022

Gratitude Chronicles

September 18, 2022

Embarking on the arduous yet rewarding journey of a PhD, I've encountered countless mentors, friends, and supporters who have played pivotal roles in shaping my academic journey. In my blog, "Gratitude Chronicles," I express heartfelt thanks to those who fueled my passion, offered guidance, and stood by me through the challenges

Arun Zachariah