Ensuring auditability and verifiability in Federated Learning (FL) is both challenging and essential to guarantee that local data remains untampered and client updates are trustworthy. Recent FL frameworks assess client contributions through a trusted central server using various client selection and aggregation techniques. However, reliance on a central server can create a single point of failure, making it vulnerable to privacy-centric attacks and limiting its ability to audit and verify client-side data contributions due to restricted access. In addition, data quality and fairness evaluations are often inadequate, failing to distinguish between high-impact contributions and those from low-quality or poisoned data. To address these challenges, we propose Federated Auditable and Verifiable Data valuation (FAVD), a privacy-preserving method that ensures auditability and verifiability of client contributions through data valuation, independent of any central authority or predefined training algorithm. FAVD utilizes shared local data density functions to construct a global density function, aligning data contributions and facilitating effective valuation prior to local model training. This proactive approach improves transparency in data valuation and ensures that only benign updates are generated, even in the presence of malicious data. Further, to mitigate privacy risks associated with sharing data density functions, we add Gaussian noise to each client’s local density function before sharing it with the server. We theoretically demonstrate the convergence, auditability, and verifiability of FAVD, along with its resilience against data poisoning threats. Our experiments on five diverse benchmarks, including three medical datasets, show that FAVD achieves significant performance gains, accurate data valuation, and fair client contributions under threat, highlighting its reliability as a trustworthy FL approach.

Anomaly detection is needed in many applications, such as video surveillance, manufacturing defect detection, and medical image analysis. However, this poses a significant challenge as anomalies can exhibit diverse patterns. Furthermore, the scarcity of labelled anomaly (defective) images complicates the use of supervised classification approaches. To address this, alternative approaches such as unsupervised or self-supervised learning are explored, effectively overcoming challenges posed by limited labelled data. While several techniques exist, there is still significant room for improvement in overall detection results, especially in handling noisy images, which is a realistic dimension often overlooked in existing methods. Therefore, we have proposed a novel framework consisting of various modules, including the Attention-Aware PDE constraint Feature Denoiser Module, Teacher, Student, and Anomaly-Attenuator. Here, we have utilized the loss function in a novel manner, which improves overall performance and ensures consistency. Additionally, we train the network in such a way that a single model would work for multiple types of objects. Finally, we considered different datasets for testing and observed that our model provides superior results, outperforming existing state-of-the-art methods for noisy and clean images.

Unambiguous customer addresses are important for the e-Commerce companies in timely and accurate delivery of shipments. In many developing countries a prescribed structure is not usually followed in practice. It is observed that the customer addresses contain additional text such as instructions to the delivery team. Further not every address has an associated geolocation information in many countries. Thus address understanding, address classification, and clustering of similar but noisy addresses are critical. In addition to last mile delivery solutions, the models help reduce buyer fraud and understanding returns. In view of noisy nature of customer addresses and non-availability of associated geolocation, the above problems are effectively solved with the help of NLP. The proposed talk traces the challenges in the Indian addresses in the context of e-commerce, solution approaches, and their extensions during the last 8 years. The talk is based on the author's own experience, his publications as well as developments in this topic across the industry over these years.

Illumination affects face recognition accuracy. When the face images are captured under uncontrolled environment, they can undergo uneven illumination due to many factors such as directional lighting, specular reflection, etc. The problem is well studied by researchers in multiple directions. We propose to solve the problem by systematically choosing a single face image as reference for computing reference probability density function and normalizing both training and test datasets of images with reference to it. We demonstrate superior performance of the proposed method and its generalization ability by conducting experiments on multiple face databases. The datasets offer challenges of varying illumination and shadows arising out of light sources placed at different angles of azimuth and elevation. We also show consistent performance of the scheme with different choices of reference images of varying illumination.

Face Detection using the Adaboost algorithm has been successfully used to detect faces in images. While the detection rate of the strong classifier trained by Adaboost is good, the false alarm rate of a single strong classifier is very high. Boosting the misclassified images during training increases the weights of the misclassified images with respect to the correctly classified images. Thus the subsequent weak classifiers are effectively trained using decreasing number of the input images. In this paper, a modification to the Adaboost method is proposed. Multiple strong classifiers based on different Haar-like feature types trained on the same set of input images are combined into a single modified-strong classifier. A comparison between the Adaboost method and the proposed method in terms of face non-face classification and face detection performance is provided. The proposed method demonstrates improved performance.

Structural Support Vector Machines (SSVMs) have recently gained wide prominence in classifying structured and complex objects like parse-trees, image segments and Part-of-Speech (POS) tags. Typical learning algorithms used in training SSVMs result in model parameters which are vectors residing in a large-dimensional feature space. Such a high-dimensional model parameter vector contains many non-zero components which often lead to slow prediction and storage issues. Hence there is a need for sparse parameter vectors which contain a very small number of non-zero components. L1-regularizer and elastic net regularizer have been traditionally used to get sparse model parameters. Though L1-regularized structural SVMs have been studied in the past, the use of elastic net regularizer for structural SVMs has not been explored yet. In this work, we formulate the elastic net SSVM and propose a sequential alternating proximal algorithm to solve the dual formulation. We compare the proposed method with existing methods for L1-regularized Structural SVMs. Experiments on large-scale benchmark datasets show that the proposed dual elastic net SSVM trained using the sequential alternating proximal algorithm scales well and results in highly sparse model parameters while achieving a comparable generalization performance. Hence the proposed sequential alternating proximal algorithm is a competitive method to achieve sparse model parameters and a comparable generalization performance when elastic net regularized Structural SVMs are used on very large datasets.

Structural Support Vector Machines (SSVMs) have become a popular tool in machine learning for predicting structured objects like parse trees, Part-of-Speech (POS) label sequences and image segments. Various efficient algorithmic techniques have been proposed for training SSVMs for large datasets. The typical SSVM formulation contains a regularizer term and a composite loss term. The loss term is usually composed of the Linear Maximum Error (LME) associated with the training examples. Other alternatives for the loss term are yet to be explored for SSVMs. We formulate a new SSVM with Linear Summed Error (LSE) loss term and propose efficient algorithms to train the new SSVM formulation using primal cutting-plane method and sequential dual coordinate descent method. Numerical experiments on benchmark datasets demonstrate that the sequential dual coordinate descent method is faster than the cutting-plane method and reaches the steady-state generalization performance faster. It is thus a useful alternative for training SSVMs when linear summed error is used.

Face Recognition remains a challenging task over two decades in spite of many advances. Face Localization and Detection forms an important step in Face Recognition systems, face tracking in video surveillance, gesture analytics, etc. Many algorithms are in use for face detection each having their relative advantages and disadvantages. The challenges of the task emanate from with image specific variations in color components, light saturation, background clutter etc. In the current work, we focus on application of Genetic Algorithms for finding optimal limits for color components as well as de-saturation that leads to highly accurate skin color detection.

Classification of large datasets is a challenging task in data mining. In the current work, we propose a novel method that compresses the data and classifies the test data directly in its compressed form. The work forms a hybrid learning approach integrating the activities of data abstraction, frequent item generation, compression, classification and use of rough sets.

Boosting is a general method for improving the accuracy of a learning algorithm. AdaBoost, short form for adaptive boosting method, consists of repeated use of a weak or a base learning algorithm to find corresponding weak hypothesis by adapting to the error rates of the individual weak hypotheses. A large, complex handwritten data is under study. A repeated use of weak learner on the huge data results in large amount of processing time. In view of this, instead of using the entire training data for learning, we propose to use only prototypes. Further, in the current work, the base learner consists of a nearest neighbour classifier that employs prototypes generated using "leader" clustering algorithm. The leader algorithm is a single pass algorithm and is linear in terms of time as well as computation complexity. The prototype set alone is used as training data. In the process of developing an algorithm, domain knowledge of the Handwritten data, which is under study, is made use of. With the fusion of clustering, prototype selection, AdaBoost and Nearest Neighbour classifier, a very high classification accuracy, which is better than reported earlier on the considered data, is obtained in less number of iterations. The procedure integrates clustering outcome in terms of prototypes with boosting.