|
BGK 596E Seminars
CRN 12338
|
| 3 |
| Title |
Encrypted Network Traffic Classification using Self-supervised Learning |
| Speaker |
Emre Can Tüzer |
| Date |
November 14th, 2025 |
| Time |
10:30 (GMT+3) |
| Microsoft Teams |
Meeting ID:
*** *** *** **
Passcode:
******
(Please, contact the secretary's office.)
|
Abstract
This paper proposes a novel two-stage self-supervised deep learning approach for encrypted network traffic classification that achieves high accuracy with only a small volume of labeled data, Critical applications including resource management, malware detection, and network provisioning are supported by network traffic classification. However, this categorization task has become difficult due to the widespread use of encrypted protocols in modern networks. Large volumes of labeled data are necessary for high accuracy in both traditional and deep learning-based supervised algorithms, but acquiring such data in actual network systems is costly, time-consuming, and requires skill. This research suggests a self-supervised learning method for classifying encrypted network traffic as a solution to this problem. With a limited amount of labeled data, the suggested approach seeks to get high accuracy. The efficacy of this method is confirmed by experiments conducted on three publicly accessible datasets: Orange'20, ISCX VPN-NonVPN, and QUIC. According to experimental results, the suggested approach performs about 3% more accurately than current state-of-the-art baseline approaches.
Furthermore, the model's transferability is remarkable; its accuracy drops by only 3% even after being pre-trained on a different dataset and refined using labeled data. For adaptive network monitoring in dynamic network conditions, this transferability is essential. These findings show that the analysis of encrypted traffic can be greatly enhanced by self-supervised representation learning, which also lessens the need for human-labeled data.
|
|
| 2 |
| Title |
LLM Agents Can Autonomously Hack Websites |
| Speaker |
Can Nair |
| Date |
November 7th, 2025 |
| Time |
10:30 (GMT+3) |
| Microsoft Teams |
Meeting ID:
*** *** *** **
Passcode:
******
(Please, contact the secretary's office.)
|
Abstract
In recent years, large language models (LLMs) have evolved to function autonomously as agents, capable of interacting with tools, reading documents, and planning multi-step tasks. This presentation covers the paper “LLM Agents Can Autonomously Hack Websites”, which investigates how LLM-based agents, under controlled experimental conditions, can autonomously perform tasks such as web reconnaissance, vulnerability discovery, and automated exploitation attempts. The experiments were conducted in a secure lab environment, following ethical guidelines, to highlight potential misuse without encouraging real-world attacks. The study shows that GPT-4-based agents can execute these tasks without prior knowledge of the vulnerabilities, demonstrating the offensive capabilities of autonomous LLM agents. These findings emphasize the dual-use potential of AI in cybersecurity and the need for careful consideration by researchers and policymakers.
|
|
| 1 |
| Title |
Approximation and Heuristic Approaches for the Metric Traveling Salesperson Problem (MTSP) |
| Speaker |
Burcu Sultan Orhan |
| Date |
October 31rd, 2025 |
| Time |
10:30 (GMT+3) |
| Microsoft Teams |
Meeting ID:
*** *** *** **
Passcode:
******
(Please, contact the secretary's office.)
|
Abstract
The Traveling Salesperson Problem (TSP) is a well-known NP-hard problem in combinatorial optimization that aims to determine the shortest possible tour visiting each city exactly once and returning to the starting point. The metric version of TSP (MTSP) assumes that the triangle inequality holds for the distance metric, i.e., d(i,j) + d(j,k) ≥ d(i,k). This assumption significantly influences the design and performance of approximation algorithms. This presentation reviews theoretical foundations and key approximation approaches for the MTSP, including the Christofides algorithm and later heuristic extensions. The focus is placed on algorithmic guarantees, complexity analysis, and the trade-off between theoretical optimality and practical efficiency.
|
|