arXiv subjects

Swarna Priya Ramu

Publications and source records attributed to Swarna Priya Ramu.

5 recordsLinked to original sources

arXiv2026

Not All Forgetting Is Equal: Retention Dynamics in Fine-Tuned Image Classifiers

Fine-tuning a pretrained classifier leaves some samples reliably learned and others cycling between correct and incorrect. Curriculum learning, data pruning and dataset cartography assume that pattern is a property of the sample, untested. We record per-sample correctness at every epoch while fine-tuning ResNet-18 and DeiT-Small on an imbalanced retinal OCT dataset and CUB-200-2011, matching samples by image identity and holding the split fixed across seeds. Per-sample retention is reproducible: cross-run Spearman correlation of the fitted decay constant is 0.37 to 0.59 over ten seeds. It is architecture-specific: two runs of one backbone agree more than two backbones on identical data (0.45 and 0.59 within against 0.30 between on OCTDL). Loss after five frozen-backbone epochs predicts a different run's decay constant at 0.29 to 0.43. The Ebbinghaus exponential does not survive: monotone decay, the one shape it can represent, is 0.1% to 0.8% of samples, and on traces that do forget mean R-squared is negative in all four configurations. A power law and a free-asymptote variant fail on the same traces: the defect is monotonicity. Across five sampling arms with matched exposure, prioritisation ratios of 2.7x to 28x, and an online variant, none of 48 comparisons against uniform sampling survives Benjamini-Hochberg correction, though three seeds detect only about four accuracy points. A stable, cheap difficulty score does not buy generalisation through sampling. Patient-grouped splitting, the remedy for a leak reaching 76% to 78% of OCT test images, moves that dataset's headline metrics by less than their run-to-run spread.

cs.LG↗

arXiv2026

VeriX-Anon: A Multi-Layered Framework for Mathematically Verifiable Outsourced Target-Driven Data Anonymization

Organisations increasingly outsource privacy-sensitive data transformations to cloud providers, yet no practical mechanism lets the data owner verify that the contracted algorithm was faithfully executed. VeriX-Anon is a multi-layered verification framework for outsourced Target-Driven k-anonymization combining three orthogonal mechanisms: deterministic verification via Merkle-style hashing of an Authenticated Decision Tree, probabilistic verification via Boundary Sentinels and exact-duplicate Twins with cryptographic identifiers, and utility-based verification via Explainable AI fingerprinting that compares SHAP value distributions before and after anonymization using the Wasserstein distance. Across seven cross-domain datasets and four cloud profiles (28 scenarios), against Lazy (drops records), Dumb (fake hash), and Approximate (valid hash) adversaries, VeriX-Anon detects 25 of 28 deviations under a fixed threshold and 27 of 28 once the threshold is calibrated per dataset, with no false alarms. No single layer achieved this alone. The XAI layer was the only mechanism that caught the Approximate adversary, succeeding on six of seven datasets and missing only a high-dimensional case where honest generalization shifts SHAP as much as the attack. Target-Driven anonymization preserved significantly more utility than blind splitting, with mean F1 gaps of +0.058 to +0.362 and Wilcoxon p <= 0.001 on six of seven datasets. Client-side verification completes under one second at one million rows. The threat model covers three empirically evaluated profiles and one theoretical Informed Attacker unable to defeat the cryptographic salt. Sentinel evasion probability ranges from near-zero to 0.82 for the most imbalanced data, which the twin layer offsets in every scenario.

cs.CR↗

arXiv2025

DeepSeqCoco: A Robust Mobile Friendly Deep Learning Model for Detection of Diseases in Cocos nucifera

Coconut tree diseases are a serious risk to agricultural yield, particularly in developing countries where conventional farming practices restrict early diagnosis and intervention. Current disease identification methods are manual, labor-intensive, and non-scalable. In response to these limitations, we come up with DeepSeqCoco, a deep learning based model for accurate and automatic disease identification from coconut tree images. The model was tested under various optimizer settings, such as SGD, Adam, and hybrid configurations, to identify the optimal balance between accuracy, minimization of loss, and computational cost. Results from experiments indicate that DeepSeqCoco can achieve as much as 99.5% accuracy (achieving up to 5% higher accuracy than existing models) with the hybrid SGD-Adam showing the lowest validation loss of 2.81%. It also shows a drop of up to 18% in training time and up to 85% in prediction time for input images. The results point out the promise of the model to improve precision agriculture through an AI-based, scalable, and efficient disease monitoring system.

cs.CV↗

arXiv2025

Evaluating Cognitive Assessment Tools:A Comparative Analysis of MMSE, RUDAS, SAGE, ADAS and MoCA for Early Dementia Detection

Early detection of dementia is very crucial to ensure treatment begins on time, however it is difficult to choose appropriate cognitive assessment tools because each test is designed differently and may not be tailored to the needs of a patient. This review compares five commonly used tests the Mini-Mental State Examination (MMSE), Rowland Universal Dementia Assessment Scale (RUDAS), Self-Administered Gerocognitive Examination (SAGE), Alzheimer's Disease Assessment Scale (ADAS), and Montreal Cognitive Assessment (MoCA). Each test has different criteria's and vary in their coverage of cognitive domains. MMSE focuses on memory and language but lacks in the evaluation of executive and visuospatial abilities. RUDAS and SAGE focus on memory, language and visual thinking while ADAS mainly targets memory, executive function and language. The MoCA is most complete as it focuses on areas like attention, memory skills, problem solving and visual skills. This review evaluates how accurate and reliable these tools are to help doctors decide the most efficient tool for diagnosis.

q-bio.QM↗

arXiv2022

Autonomous Vehicles in 5G and Beyond: A Survey

Fifth Generation (5G) technology is an emerging and fast adopting technology which is being utilized in most of the novel applications that require highly reliable low-latency communications. It has the capability to provide greater coverage, better access, and best suited for high density networks. Having all these benefits, it clearly implies that 5G could be used to satisfy the requirements of Autonomous vehicles. Automated driving Vehicles and systems are developed with a promise to provide comfort, safe and efficient drive reducing the risk of life. But, recently there are fatalities due to these autonomous vehicles and systems. This is due to the lack of robust state-of-art which has to be improved further. With the advent of 5G technology and rise of autonomous vehicles (AVs), road safety is going to get more secure with less human errors. However, integration of 5G and AV is still at its infant stage with several research challenges that needs to be addressed. This survey first starts with a discussion on the current advancements in AVs, automation levels, enabling technologies and 5G requirements. Then, we focus on the emerging techniques required for integrating 5G technology with AVs, impact of 5G and B5G technologies on AVs along with security concerns in AVs. The paper also provides a comprehensive survey of recent developments in terms of standardisation activities on 5G autonomous vehicle technology and current projects. The article is finally concluded with lessons learnt, future research directions and challenges.

cs.NI↗