Suleman Khan

DL algorithms have shown wonders in many NLP tasks but are vulnerable to adversarial attacks. Most mitigation techniques proposed to date are supervised, relying on adversarial retraining to improve the robustness.

We introduce an unsupervised detection methodology for detecting adversarial inputs to NLP classifiers. We note that minimally perturbing an input to change a model's output, a significant strength of adversarial attacks, is a weakness that leaves unique statistical marks reflected in the cumulative contribution scores of the input. Particularly, we show that the cumulative contribution score, called the CF-score of adversarial inputs, is generally greater than that of the clean inputs. We thus propose Con-Detect, a contribution-based detection method for detecting adversarial attacks against NLP classifiers. Con-Detect can be deployed with any classifier without having to retrain it. We show that it can reduce the attack success rate (ASR) of different attacks from 100% to as low as 0% for the best cases and =70% for the worst case. Even in the worst case, we note a 100% increase in the required number of queries and a 50% increase in the number of words perturbed, suggesting that Con-Detect is hard to evade.

While the technique of deep neural networks (DNNs) has been instrumental in achieving state-of-the-art results for various natural language processing (NLP) tasks, recent works have shown that the decisions made by DNNs cannot always be trusted. Recently proposed explainable artificial intelligence (XAI) methods are open to attack and can be manipulated in both white-box gradient-based and black-box perturbation-base scenarios.

We proposed first-of-its-kind Tamp-X, a novel attack that tampers the activations of robust NLP classifiers forcing the state-of-the-art white-box and black-box XAI methods to generate misrepresented explanations. Through extensive experimentation, we show that the explanations generated for the tampered classifiers are unreliable and significantly disagree with those generated for the untampered classifiers, even though the output decisions of tampered and untampered classifiers are almost always the same. Additionally, we study the adversarial robustness of the tampered NLP classifiers and find out that the tampered classifiers, which are harder to explain for the XAI methods, are also harder to attack by adversarial attackers.

With the hyperconnectivity and ubiquity of the Internet, the fake news problem now presents a greater threat than ever before. One promising solution for countering this threat is to leverage deep learning (DL)-based text classification methods for fake-news detection. However, since such methods are vulnerable to adversarial attacks, the integrity and security of DL-based fake news classifiers are under question. We evaluate the performance of fake-news detectors under various configurations under black-box settings. In particular, we investigate the robustness of four different DL architectural choices, such as MLP, CNN, RNN, and a recently proposed Hybrid CNN-RNN trained on three different state-of-the-art datasets, under other adversarial attacks like Text Bugger, Text Fooler, PWWS, and Deep Word Bug implemented using TextAttack.

Additionally, we explore how changing the detector complexity, the input sequence length, and the training loss affects the robustness of the learned model. Our experiments suggest that RNNs are robust compared to other architectures and our evaluations provide vital insights to robustify fake-news detectors against adversarial attacks.

NASA is moving its data to the cloud, and machine learning (ML) and artificial intelligence (AI) can offer an innovative means to analyze and use this massive archive of free and open data.

We developed a mobile application using ML/AI techniques that allows users to input short text phrases, match that input to NASA science data or imagery, and display the results for the user creatively and artistically.

Humans have seven distinct facial emotions. Facial expression recognition algorithms have applications in healthcare, entertainment, criminal justice, and more.

Deep learning algorithms are efficient for facial expression classification, but these algorithms demand a high amount of data. Domain Adaptation can be used to address the lack of sufficient data. Right Now, we don't have much data on Pakistani facial expressions. In this Project, we created data set of Pakistani facial expressions and used domain adaptation to develop an efficient facial expression algorithm of Pakistani faces. We have achieved 58% accuracy with a baseline of 32 %.

OpenAI's GPT-3 model is a very powerful model which contains 175 billion parameters and was trained on 45TB of text from various data sets.

I developed the project to demonstrate the capabilities of state-of-the-art NLP models. This demonstrates how we can use GPT Neo and GPT-3 to write blogs.

The Bookstore Application is an eCommerce platform built using the Django web framework, SQL databases, and Postgres as the database engine. The platform allows users to browse and purchase books from a local bookstore.

The application features a user-friendly interface that allows users to easily search for books by title, author, or genre. Users can add books to their shopping cart and proceed to checkout, where they can enter their payment information and complete their purchases.

The back end of the application is built using Django, which allows for efficient data management and user authentication. The application uses SQL and Postgres to store and manage user and product data, ensuring data security and scalability.

Overall, the Bookstore Application is a versatile and user-friendly platform that enables the local bookstore to sell their books online, providing customers with a convenient way to shop and support their favorite local bookstore.