Abstrak The transition toward sustainable energy has become a strategic priority for Indonesia, particularly through the implementation of bioethanol-blended fuel policies. However, public perception toward this policy remains diverse and dynamic, especially as expressed on social media platforms. This study aims to analyze public sentiment regarding the implementation of bioethanol-blended fuel (E10) policies on X (Twitter) and to compare the performance of traditional machine learning and Transformer-based models in sentiment classification. This research adopts a quantitative experimental approach using Natural Language Processing (NLP) techniques. A total of 2,501 tweets were collected through web crawling and processed using a Dual Pipeline Preprocessing approach. Sentiment labeling was conducted using the VADER method with manual validation. Two classification models were implemented, namely Support Vector Machine (SVM) as the baseline model and IndoBERT as the Transformer-based model. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics. The results indicate that the IndoBERT model outperforms SVM, achieving an accuracy of 81.64% and an F1-score of 81.39%, compared to SVM with an accuracy of 72.46% and an F1-score of 72.02%. The performance improvement of 9.18% demonstrates the superiority of Transformer-based models in capturing contextual semantics in unstructured social media text. In addition, sentiment analysis results reveal that public opinion is predominantly positive toward the policy, although concerns regarding technical and economic aspects remain. This study contributes by providing empirical insights into public perception of energy policy and demonstrating the effectiveness of Transformer-based models for sentiment analysis in the Indonesian language context

This research is in the field of cybersecurity and artificial intelligence, focusing on developing a real-time attack detection system for MariaDB databases. The research was motivated by the increasing threats of SQL injection (SQLi), brute-force attacks, and data exfiltration, which are increasingly difficult to detect using conventional rule-based methods. The objective of the research was to develop a hybrid architecture based on Long Short-Term Memory (LSTM) and Isolation Forest to improve detection accuracy while reducing false positives in modern database environments. The research method employed a Research and Development (R&D) approach through three main modules. The first module applies a bidirectional LSTM to detect SQLi based on query sequence analysis. The second module uses Isolation Forest and Rotated Isolation Forest to detect brute-force attacks through access behavior analysis. The third module applies Isolation Forest to detect data exfiltration based on traffic patterns and data transfer behavior. The entire dataset underwent preprocessing, feature engineering, tokenization, normalization, and performance evaluation using confusion matrices, precision, recall, F1-score, and AUC. The results show that the Bi-LSTM model achieved 99.99% accuracy in detecting SQLi. In brute-force detection, the standard Isolation Forest provided the best performance with a recall of 99.94% and an F1-score of 99.61%. Meanwhile, the data exfiltration module achieved a 100% detection rate on simulated exfiltration traffic with a combined accuracy of 94.92%. This study proves that the hybrid LSTM–Isolation Forest architecture is capable of providing accurate, adaptive detection and is feasible to be implemented as a next-generation MariaDB database security system.

Network forensic investigations rely heavily on the integrity and traceability of Packet Capture (PCAP) files as primary digital evidence. Digital Forensic Research Workshop (DFRWS) implementations commonly employ centralized preservation mechanisms that remain vulnerable to unauthorized modification and provide limited provenance transparency. To address these limitations, this study proposes a blockchain-based preservation framework integrated into the preservation phase of the DFRWS model. The framework combines SHA-256 cryptographic hashing for integrity verification, blockchain-based provenance logging, and distributed ledger validation while maintaining off-chain evidence storage. Unlike many existing blockchain-based forensic frameworks that primarily emphasize provenance recording and chain-of-custody management, this study evaluates evidence preservation through an integrated validation approach consisting of controlled tampering simulation, cryptographic sensitivity analysis, and preservation latency measurement. Experimental evaluation using PCAP datasets representing attack and baseline traffic conditions demonstrated that unauthorized evidence modification was successfully detected through hash inconsistencies. Avalanche Effect analysis produced a value of 50.39%, confirming the strong cryptographic sensitivity of the SHA-256 mechanism to minimal data alteration. While SHA-256 enables reliable tampering detection, the integrated blockchain architecture provides tamper-resistant provenance recording, chain-of-custody traceability, and distributed verification of evidence integrity. The framework achieved an average preservation latency of 2.057 seconds within the experimental environment, providing preliminary evidence of feasibility for blockchain-assisted forensic logging under controlled conditions. Although no direct comparison with alternative preservation approaches was conducted, the findings provide a proof-of-concept validation and contribute empirical evidence regarding the potential of blockchain-supported provenance management to enhance trustworthiness and integrity assurance in network forensic workflows.