Research Statement

Research Vision

Most successful cyberattacks still begin with a person being persuaded to click, reply, or trust. My research builds AI systems that recognize this persuasion and stop it before it reaches the user. I want these systems to stay accurate when attackers adapt and to explain their decisions to the analysts who must act on them. My work sits where natural language processing, deep learning, and security engineering meet, and it draws on more than 23 years across academia, research, and industry.

Current Research: Behavior-Aware Phishing Detection

The problem

Phishing remains one of the most widespread and costly forms of cybercrime. Attackers now use generative AI, look-alike domains, and personalized social engineering to slip past URL blacklists and signature-based filters. Current defenses sit either at the network perimeter or on the endpoint, and both struggle with zero-day campaigns that carry no known signature. By the time a phishing message reaches the inbox, the user is already exposed.

My approach

As Postdoctoral Fellow at SR University (2025–present), under the supervision of Dr. V. Venkataramana, I am developing a detection layer positioned between OSI Layers 5 and 6. It intercepts text communications, including email, SMS, and URLs, and analyzes them before delivery, within a target budget of 500 ms per message.

The central idea is that phishing works by pulling predictable psychological levers, and those levers can be measured. Building on Cialdini’s principles of persuasion, the system represents each message as a four-dimensional behavioral vector of urgency, fear, authority, and reward. Rule-based keyword and pattern scoring captures explicit cues, and a fine-tuned BERT model captures emotional manipulation that only appears in context. As a result, the system can flag an attack even when the link or attachment itself is not yet known to be malicious.

Results

Evaluated on 2023–2024 data (14,700 PhishTank URLs, 8,500 Gmail phishing emails, and 4,000 SMS messages in four languages), the framework achieves:

• Overall performance: 95% accuracy, 0.98 mean AUC-ROC, and 0.96 F1-score, ahead of RoBERTa-Large (94.0%), ALBERT (92.9%), LSTM (92%), XGBoost+BERT (91.6%), and Phishpedia (88.3%)
• Multilingual detection: 93.0% mean accuracy across English, Arabic, Mandarin, and Hindi with a fine-tuned multilingual BERT
• Adversarial robustness: an accuracy drop of only 2.9% under character-perturbation attacks with adversarial training, against 12% for the baseline
• Explainability: attention visualization and SHAP token attributions that tie each decision to the behavioral cue behind it

These results close the two gaps left by an earlier version of the work, which reached only 82% on non-English SMS and lost 12% accuracy under text perturbation.

Why it matters

Most phishing research evaluates English text only, while users in the Gulf and South Asia receive scams in Arabic, Hindi, and other languages, including banking and UPI fraud. Multilingual detection therefore addresses a real gap in the regions I have worked in. The design also takes privacy seriously: it anonymizes data, excludes encrypted traffic from content inspection, and balances classes to reduce bias.

Research Foundations

My current work grows out of three earlier threads, each of which contributed methods I still use.

AI-driven intrusion detection and security architecture

Much of my security work has addressed intrusion detection under realistic constraints. I have tackled class imbalance with GAN-based augmentation [22] and with SMOTE and a hybrid SVNN model [19], generated synthetic security data with GANs [21], and detected attacks in MANETs [25] and wireless mesh networks [18]. Most recently, I developed explainable, attention-based intrusion detection for encrypted 5G traffic [33]. On the architecture side, I have contributed a zero-trust supply-chain framework for AI systems built on the Model Context Protocol [32], a multi-layered banking security framework [27], and an analysis of how consumer IoT devices are exposed to phishing and ransomware [16].

Natural language processing for online safety

My work on sarcasm detection for cyberspace safety, published in Scientific Reports [24], deals with a problem closely related to phishing: identifying intent that the text does not state literally. This experience with context-dependent language fed directly into the behavioral modeling behind my phishing research.

Deep learning for medical diagnosis

My Ph.D. at KIIT University (2017–2024) developed deep learning methods for detecting COVID-19 from 2D and 3D medical images, including CNN-GRU hybrids [8] and graph convolutional networks [10]. I later extended this work to multiple lung diseases using both CT scans and cough sounds [12]. This period gave me a standard I apply in every high-stakes detection problem: a model must be accurate, robust to shifts in data, and able to explain itself. The same standard runs through my explainable-AI work in healthcare [26], [31] and in security [33].

Research Impact

My work includes 33 publications (25 Scopus-indexed; 183 citations and an h-index of 7 on Google Scholar), five patent applications in the United Kingdom and India, and four authored books. It received the 2nd Best Paper award at UNESCO 2025. I also supervise two Ph.D. scholars at SR University on liquid neural networks for land-cover classification and on hyperspectral crop monitoring, which shows I can guide doctoral research beyond my own specialty.

Future Directions: Securing UPI Payments

Over the next three to five years, I will focus my program on India’s Unified Payments Interface (UPI), the world’s largest real-time retail payment system. My earlier security analysis of UPI across its device, application, network, and settlement layers showed where the ecosystem is exposed. My phishing research gives me tools to address one of the most common ways it is attacked: persuading users to approve payments themselves. I plan to pursue five directions:

• Detecting UPI social-engineering fraud: I will adapt the behavioral vector to UPI scams such as fake collect requests, fraudulent payment links and QR codes, and fake refund or KYC messages delivered by SMS, WhatsApp, and voice calls, and harden the models against scammers who change their wording.
• Indian-language coverage: UPI users receive scams in many languages, yet my current Hindi subset contains only 250 messages. I aim to build open, labeled fraud datasets for Hindi, Telugu, Tamil, and other Indian languages.
• Real-time transaction anomaly detection: I will combine message-level signals with transaction patterns to flag suspicious payments before they settle, using models efficient enough to run at national transaction scale.
• Auditable fraud decisions: I will develop explainable-AI methods that make every automated fraud flag traceable, so that banks, payment providers, and regulators can audit and contest decisions.
• Beyond UPI: I will turn the vulnerability-class analysis into concrete mitigations and test how the methodology transfers to other real-time payment systems, including those in the Gulf region.

I intend to build a research group of Ph.D., master’s, and undergraduate students, pursue external funding, and partner with banks, payment apps, and telecom operators for realistic data and deployment settings. I welcome collaboration on financial fraud detection, adversarial machine learning, and multilingual NLP.

Publications

Indexing and quartiles: SCIE/ESCI and JCR from Clarivate Web of Science; SJR from SCImago (Scopus); latest available year. Citation numbers in the text above link to the entries below.

Current Manuscript

1. [M1]
   R. Dhumpati and V. R. Veeramsetty, “Robust Multilingual Phishing Detection Using Behavioral Psychology and Adversarially-Trained Transformers.”