Institutional Repository

U n i v e r s i t y of P e r a d e n i y a

The Institutional Repository of the University of Peradeniya is the University's digital gateway to scholarly knowledge and research excellence. Established as the next phase of the Digital Library initiative launched by the University Library in 2011, it preserves, showcases, and provides open access to the intellectual output of the University community.

Through a rich collection of research articles, conference proceedings, theses, dissertations, and other scholarly works, the repository enhances the visibility, accessibility, and long-term preservation of knowledge, extending the global reach and impact of the University of Peradeniya’s academic and research contributions.

Photo by @ITD,Main Library

Communities in DSpace

Select a community to browse its collections.

Now showing 1 - 5 of 5

Recent Submissions

MESH: a multi-stage benchmark for mockery, extremism, sarcasm, and hate speech jailbreak vulnerabilities in open-source LLMs
(University of Peradeniya, Sri Lanka, 2026-08-20) Mithushika, S.; Sweshthika, S.; Kayalvizhy, T.
Jailbreak attacks enable large language models (LLMs) to generate harmful, unethical, or otherwise objectionable content, posing a significant safety concern. Existing benchmarks relying on first-generation static templates yield artificially suppressed attack success rates that poorly discriminate model vulnerability. This study introduces a multi-stage adaptive benchmarking framework spanning four harm categories, termed MESH (Mockery, Extremism, Sarcasm, Hate Speech), across three semantic targets (Social Norms, Protected Groups, Democratic Institutions). The dataset was constructed synthetically, with no external dataset or pre-existing prompt repository used, via a three-dimensional Cartesian product across the four harm categories, three semantic targets, and twelve attack templates, yielding 288 adversarial prompts through combinatorial template-target pairing, supplemented by 22 benign controls (10 clearly-safe, 12 borderline-safe). The twelve attack templates span two generations — six established Gen-1 templates (Role-play Wrapper, CoT Leakage, Adversarial Persona, Virtualization, Academic Framing, Payload Splitting) and six novel Gen-2 templates introduced by this work (Contextual Camouflage, Gradual Escalation, Semantic Obfuscation, Translation Attack, Moral Inversion, Logical Inversion), each assigned an a priori difficulty score (0.50-0.80) to enable subsequent difficulty-banded analysis. This generational and difficulty-graded design was deliberately chosen to test whether attack sophistication predicts effectiveness. Responses generated by LLaMA 3.2-3B were evaluated using a dual-model ensemble judge (Mistral-7B and Gemma-7B), selected for heterogeneous training origins to reduce correlated judgment bias; LLaMA 3.2-3B was restricted to a meta-judge role only, to avoid self-evaluation bias on its own outputs. A two-stage judging protocol separated harmfulness detection from four-class labeling to improve labeling consistency, with disagreements resolved via a harm-union policy and a benign-prompt guard preventing false-positive Bypass labels. Judge outputs were validated against gold labels from three independent human annotators (Fleiss' κ = 0.7573, substantial agreement). The pipeline achieved an overall attack success rate (ASR) of 41.11% (human gold labels), a 53.31% refusal rate on adversarial prompts, and a 95.45% safe-response rate with only 4.55% over-refusal on benign controls. CoT Leakage (79.17%) and Logical Inversion (70.83%) were the most effective attack types, while Mockery achieved the highest category-level ASR (56.94%). The ensemble judges reached 77.74% inter-model agreement with only a 0.49% ASR delta against human labels, supporting ensemble judging as a reliable substitute for large-scale human annotation. A notable difficulty- band paradox emerged: easy prompts (55.56% ASR) outperformed hard prompts (22.92% ASR), indicating that current safety alignment relies more on surface pattern recognition than genuine semantic understanding. A multi-criteria curation process, jointly filtering on confidence (≥0.65), inter-annotator agreement, and difficulty (≥0.70), yielded 105 high-confidence reference prompts suitable for Supervised Fine- Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF).
Machine learning-based approach for classifying damaged characters in early Brahmi Sri Lankan inscriptions
(University of Peradeniya, Sri Lanka, 2026-08-20) Guruge, G. C. J. I.; Thilakarathna, U. D. R. S.; Dissanayake, M. B.; Pinidiyaarachchi, A. J.; Malsiri, W. I.; Niranjan, M.
Early Brahmi Inscriptional character recognition for epigraphy is a computer vision task challenged by data scarcity, in which surface weathering, biological growth, and physical erosion progressively degrade the visual representation of the character. Although prior benchmarks have established performance ceilings on intact inscriptional characters, no study has quantified how those methods generalise to glyphs of varying damage severity. This work addresses that gap through a controlled external-validation protocol applied to three top-performing classical methods identified as promising in earlier work: support vector machines (SVM) with Radial Basis Function (RBF) kernel, k-nearest neighbours (k-NN), and multi-layer perceptrons (MLP). A locally collected 545-image, six-class dataset of Early Brahmi Inscriptional characters extracted from field-study was stratified into three curator-assigned damage tiers: undamaged (n=388), mildly damaged (n=61), and severely damaged (n=96), with samples from the latter two held out for testing. Histogram-of-Oriented-Gradients (HOG) features were extracted at 64×64 resolution following Contrast Limited Adaptive Histogram Equalisation (CLAHE) preprocessing. A two-stage 5-fold cross-validation protocol was applied: hyperparameters were selected by inner-loop grid search on the training pool, then evaluated through an outer 5-fold loop in which each fold's model was tested on a held-out 20% of the undamaged, mildly-damaged, and severely-damaged pools. SVM-RBF achieved the highest accuracy at every damage level (Undamaged: 0.730 ± 0.038; Mildly: 0.685 ± 0.038; Severely: 0.496 ± 0.018), with non-overlapping error bars to k-NN (Severely: 0.406 ± 0.011) and MLP (Severely: 0.408 ± 0.040). All the mildly and severely damaged image classifications were closely evaluated and validated by the domain experts. All three methods exhibited a non-linear performance cliff, averaging 17.4 percentage points between mild and severe damage, and one specific class became effectively unclassifiable across every method (F1 < 0.10), empirical evidence for a feature-space-level limit on classical methods that motivates the use of learned representations and image restoration as complementary approaches to closing the damage-induced accuracy gap.
Impact of progressive corrosion on nonlinear damping behavior in reinforced concrete elements
(University of Peradeniya, Sri Lanka, 2026-08-20) Dissanayake, G. B.; Dammika, A. J.; Bandara, C. S.; Jayasinghe, J. A. S. C.
Corrosion of reinforcement in reinforced concrete (RC) structures is a well-known cause of deterioration, which can severely affect the structural integrity and safety. Traditional methods for assessing corrosion damage, such as visual inspection and localized testing are widely used, but they often fail to detect early-stage corrosion, particularly when surface damage is not visible. Therefore, potential of vibration-based damage detection methods to detect corrosion damage has been the interest of recent research and modal damping has been drawn attention due to its pronounced sensitivity to minor damage. Therefore, this study investigates the effect of progressive corrosion on the modal vibration behavior of RC elements, focusing on how changes in damping characteristics can offer valuable insights into the early stages of corrosion damage. This research proposes a novel approach for non-destructive monitoring of corrosion-induced degradation in RC structures by analyzing the nonlinear changes in damping behavior. In this study, T-shaped RC specimens with concrete grades C25, C30, C40, and C50 were subjected to progressive corrosion through the impressed current method. The corrosion levels were controlled to induce localized multiple damage stages, allowing the investigation of its effect on the modal damping response. Free vibration tests were conducted after each corrosion level, and the time history data were processed using Variational Mode Decomposition to isolate mono-frequency modal vibration response. The study used both Decay Curve Method for overall damping analysis and the Instantaneous Logarithmic Decrement method to obtain insights into the nonlinear damping behavior and its sensitivity to corrosion damage. The results revealed that corrosion-induced changes in modal damping behavior were more complex than previously understood. The magnitude and time domain variation of the damping response were strongly influenced by the progression of corrosion. As corrosion develops, the damping behavior changed increasingly nonlinear particularly during its early stages. This nonlinear behavior was observed to be diminished beyond a certain threshold, at which state corrosion-induced cracks become sufficiently wide that interfacial friction between crack surfaces is no longer effective in energy dissipation. The study demonstrated the sensitivity of nonlinear damping to early corrosion damage, particularly in high strength concrete. It was observed in the experiment that even 2% mass loss due to corrosion was detected through proposed nonlinear damping analysis method which is not feasible in conventional visual inspection methods. These findings provide a new insight into the potential of nonlinear modal damping as a sensitive and reliable vibration-based damage indicator for detecting the corrosion-induced damage in RC structures at its initiation.
Hybrid deep learning-based classification of type-1 and type-2 code clones across multiple programming languages
(University of Peradeniya, Sri Lanka, 2026-08-20) Prabodhani, H. K. G. S.; Vasanthapriyan, S.; Rubasinghe, T. D.
Code clone detection is a crucial software engineering task that supports software maintenance, code reuse, and defect prevention by identifying duplicated or highly similar source code fragments. However, detecting Type-1 and Type-2 code clones across multiple programming languages remains challenging due to variations in syntax, lexical characteristics, and structural representations. Existing clone detection techniques often exhibit limited cross-language adaptability as they are primarily designed for individual programming languages or depend on a single form of code representation. This research proposes a hybrid deep learning framework that integrates lexical and structural code representations for effective Type-1 and Type-2 code clone classification across Java, Python, and C++ programming languages. A quantitative empirical evaluation approach is adopted using benchmark datasets, including BigCloneBench, CodeXGLUE, and CodeSearchNet. The source code samples undergo preprocessing steps, including tokenization, normalization, identifier abstraction, class balancing, and duplicate removal, to improve data consistency and minimize potential data leakage. The datasets are partitioned into training, validation, and testing subsets using a consistent 70:15:15 ratio to ensure fair and reliable model evaluation. The proposed Hybrid CNN-BiLSTM-Attention framework integrates lexical token sequences with Abstract Syntax Tree (AST)-based structural representations. CNN layers extract local syntactic patterns, BiLSTM networks capture contextual dependencies, the Attention mechanism emphasizes significant code regions, and AST features provide complementary structural information. The framework is evaluated against baseline deep learning models, including CNN, LSTM, and Transformer-based approaches, using accuracy, precision, recall, F1-score, ROC-AUC, confusion matrix, and cross-language evaluation metrics. The experimental results demonstrate that the proposed framework achieves 92% accuracy, outperforming the evaluated baseline models. Cross-language evaluation, performed by training on two programming languages and testing on an unseen third language, achieves accuracies of 83.2%, 84.5%, and 82.8% for C++, Java, and Python test scenarios, respectively. These results demonstrate strong multilingual generalization capability and confirm that combining lexical and AST-based structural representations improves Type-1 and Type-2 code clone detection performance. The proposed framework provides a scalable and effective solution for automated multilingual code clone classification. This research contributes a hybrid representation-based approach that enhances cross-language code analysis, reduces code duplication, and supports intelligent software engineering practices.
Hybrid deep learning framework for unsupervised anomaly detection in DOGE/USD cryptocurrency price movements
(University of Peradeniya, Sri Lanka, 2026-08-20) Kaushalya, H. W. S.; Napagoda, N. A. D. N.
Cryptocurrency markets play a significant role in global finance, impacting international trade. Due to its volatile and nonlinear behavior, anomaly detection has become a critical task for financial risk analysis. Beyond the traditional statistical approaches, this study explores unsupervised machine learning and deep learning techniques to identify abnormal movements in cryptocurrency price data. In this study, the main objective is to develop and evaluate a hybrid anomaly detection framework that combines Isolation Forest and Autoencoder models to detect abnormal price movements in Dogecoin US Dollars Price data. It further compares the effectiveness of individual models with the proposed hybrid approach using quantitative evaluation metrics. The proposed methodology utilizes historical DOGE/USD OHLC (Open, High, Low, Close) data, covering the period from 2017 to 2025, which are normalized using Min-Max scaling. An Isolation Forest model is trained to capture global statistical deviations, while an Autoencoder neural network is trained to capture structural anomalies. A hybrid framework is constructed by combining Isolation Forest and Autoencoder. Since labeled data is unavailable, evaluation is performed using unsupervised metrics. Isolation Forest identified 44 anomalies but showed weaker separation, while the Autoencoder detected 27 anomalies with stronger sensitivity to extreme price changes. The hybrid model detected 38 anomalies and achieved the highest separation score of 0.486, compared to Isolation Forest (-0.198) and Autoencoder (0.046). The results indicate that nonlinear pattern-based anomalies dominate the dataset, and the combination of both models provides a more reliable representation of market behavior. Overall, the proposed hybrid framework provides more reliable and consistent anomaly detection compared to standalone models by integrating statistical and deep learning approaches for cryptocurrency markets. This study contributes to the existing body of knowledge by demonstrating that the integration of Isolation Forest and Autoencoder improves the effectiveness of unsupervised anomaly detection for cryptocurrency price movements.