Presented by: Zhuo Chen

Date: Thursday, August 1, 2024
Time: 10:00 am
Place: Zoom - Please see below

Join Zoom Meeting

https://uvic.zoom.us/j/81148722115?pwd=tnvRdGxgo3VvSSmwrlVaGJcrylZaWQ.1

Meeting ID: 811 4872 2115

Password: 105375

One tap mobile

+17789072071,,81148722115# Canada

+16475580588,,81148722115# Canada

Dial by your location

        +1 778 907 2071 Canada

        +1 647 558 0588 Canada

Meeting ID: 811 4872 2115

Find your local number: https://uvic.zoom.us/u/kJq94LcBg

Abstract: 

As Internet of Things (IoT) devices, drones are among the most popular unmanned aerial vehicles (UAVs), equipped with multiple sensors, cameras, and communication systems. These features expose them to potential vulnerabilities exploitable by hackers. making it crucial to explore these vulnerabilities and implement effective anomaly detection while operating UAVs. This study investigates a DJI Edu Tello drone to comprehensively assess its vulnerabilities and develop anomaly detection mechanisms using different unsupervised machine learning techniques. Two types of data were collected: benign data from legitimate actions and attack data comprising nine types of attacks. Feature extraction and engineering were performed based on scripts from the Canadian Institute for Cybersecurity (CIC), which were modified to suit the specific needs of this project. The modifications aimed to improve the robustness of the detector by removing and modifying existing features and introducing new measurements to represent the captured packets. The anomaly detector was formulated after comparing three unsupervised machine learning algorithms: Isolation Forest, Local Outlier Factor (LOF), and Elliptic Envelope, through extensive performance evaluations and analyses. The study demonstrated the effectiveness of these algorithms in detecting anomalies and enhancing the security of drones. The findings also highlight the critical role of robust feature engineering and careful algorithm selection in developing a reliable anomaly detection system for UAVs.

Presented by: Brosnan Yuen

Date: Thursday, August 1, 2024
Time: 1:00 pm
Place: Zoom, link below.

Location: Zoom Meeting ID: 850 8650 1062   , Zoom Password: 284068

Abstract: This research proposes a universal activation function (UAF) that achieves near optimal performance in quantification, classification, and reinforcement learning (RL) problems. For any given problem, the gradient descent algorithms are able to evolve the UAF to a suitable activation function by tuning the UAF’s parameters. For the CIFAR-10 classification using the VGG-8 neural network, the UAF converges to the Mish like activation function, which has near optimal performance F1 = 0.902±0.004 when compared to other activation functions. In the graph convolutional neural network on the CORA dataset, the UAF evolves to the identity function and obtains F1 = 0.835 ± 0.008. For the quantification of simulated 9-gas mixtures in 30 dB signal-to-noise ratio (SNR) environments, the UAF converges to the identity function, which has near optimal root-mean-square
error of 0.489 ± 0.003 μM. In the ZINC molecular solubility
quantification using graph neural networks, the UAF morphs to
a LeakyReLU/Sigmoid hybrid and achieves RMSE=0.47 ± 0.04.
For the BipedalWalker-v2 RL dataset, the UAF achieves the 250
reward in 961 ± 193 epochs with a brand-new activation function, which gives the fastest convergence rate among the activation functions.

Presented by: Soroush Danaeifard

Date: Monday, August 5, 2024
Time: 9:00 am
Place: Remote Via Zoom

Title: A Machine Learning Framework for Malware Triage 

Speaker: Soroush Danaeifard 

Supervisors: Dr. Issa Traore 

 Date & Time: Aug 5, 2024, Monday, 09:00 am.

 Pacific Time (Canada and US)

 Location: Remote via Zoom 

 Join Zoom Meeting

https://uvic.zoom.us/j/7309466548?pwd=TVo5S201TEtVOENMOXJIQkxMdmF0UT09

Meeting ID: 730 946 6548 

Password: 645466 

Note: Please log in to Zoom via SSO and your UVic Netlink ID 

 ABSTRACT

 Every day, thousands of new malicious software emerge globally, posing threats to consumer devices, stealing private data, or inducing financial losses. The increasing number and sophistication of malware threats underscores the need for effective and efficient malware detection and triage schemes. Malware triage is a process used by cybersecurity professionals to quickly assess, prioritize, and respond to malware incidents. Effective malware triage requires a combination of automated tools, skilled personnel, and well-defined procedures to respond to malware incidents quickly and accurately, minimizing damage and recovery time. An essential aspect of the triage is the categorization and prioritization task which enables determining malware type. Automating such process helps save valuable time for cybersecurity analysts and investigators.

Presented by: Shahin Honari

Date: Tuesday, August 6, 2024
Time: 11:00 am
Place: Zoom, link below.

Abstract: 

This work investigates the fabrication and characterization of Silica on Silicon (SoS) microcavities, for potential comb generation and sensing applications. 
In this thesis, novel approaches were introduced on silica-on-silicon microdisks to remove some of the inherent hurdles, namely mitigating the rough surface through chemo-mechanical polishing and suppressing buckling-induced mechanical instability by introducing a rib disk structure. By addressing these challenges, we unleash the enormous potentials of ultra-high Q silica microdisks in microcomb generation in visible regime and sensing. Furthermore, a multi-transverse mode dense comb generation scheme on this platform is investigated.

Shahin Honari is inviting you to a scheduled Zoom meeting.

Topic: Graduate Student Seminar-Shahin Honari

Time: Aug 6, 2024 11:00 AM Vancouver

Join Zoom Meeting

https://uvic.zoom.us/j/85428594992?pwd=lmUZdghLb3MZbGRER9XSLgtUwLhbu6.1

Meeting ID: 854 2859 4992

Password: 972495

One tap mobile

+16475580588,,85428594992#,,,,0#,,972495# Canada

+17789072071,,85428594992#,,,,0#,,972495# Canada

Dial by your location

        +1 647 558 0588 Canada

        +1 778 907 2071 Canada

Meeting ID: 854 2859 4992

Password: 972495

Find your local number: https://uvic.zoom.us/u/kdVqOprOmT

Presented by: Zahra Nikdel

Date: Friday, August 9, 2024
Time: 1:00 pm
Place: Zoom, link below.

Join Zoom Meeting

https://uvic.zoom.us/j/89535343960?pwd=sfirnTWfmcHGpH0bkhoaYxakRXrgr7.1

Meeting ID: 895 3534 3960

Password: 091121

One tap mobile

+16475580588,,89535343960#,,,,0#,,091121# Canada

+17789072071,,89535343960#,,,,0#,,091121# Canada

Dial by your location

        +1 647 558 0588 Canada

        +1 778 907 2071 Canada

Meeting ID: 895 3534 3960

Password: 091121

Find your local number: https://uvic.zoom.us/u/kk9mOuDix

Note: Please log in to Zoom via SSO and your UVic Netlink ID 

ABSTRACT

In this seminar, we explore the concept of ergodicity in software-centric systems, ranging from simple embedded systems to complex distributed ones in commercial clouds. Modern software technologies increasingly utilize control theory and machine learning to address challenges such as performance prediction and reliability analysis. These methodologies presuppose that the systems in question are both ergodic and stationary, enabling future behavior predictions based on historical data. We begin by modeling stochastic runtime performance measures as dynamical systems, focusing on macro-level observables like average response times. The seminar will cover formal definitions of statistical stationarity and ergodic theory, highlighting Birkhoff's Ergodic Theorem (BET) and Peter Walters' conditions for testing ergodicity. The discussion then shifts to the applicability of these theories to large-scale distributed systems (LDSS). We identify that while some conditions are easily applicable, others present significant challenges. Factors such as fair scheduling, server overloading, queue drops, and reliable protocols are analyzed for their impact on maintaining ergodicity. We discuss strategies such as the application of RTOS, careful server utilization management, and preventing queue overflows to maintain measure invariance and ergodicity in deployed systems. The insights gained from this investigation lead to the development of four software engineering design rules essential for ensuring BET-compliance in software systems. This seminar provides a deep understanding of ergodicity in software systems, equipping attendees with the knowledge to address related challenges in their own work.

Presented by: William Briguglio

Date: Wednesday, August 14, 2024
Time: 11:00 am
Place: Zoom, link below.

Note: Please log in to Zoom via SSO and your UVic Netlink ID.

Join Zoom Meeting:

https://uvic.zoom.us/j/84368416374

Meeting ID: 843 6841 6374

One tap mobile

+16475580588,,84368416374# Canada

+17789072071,,84368416374# Canada

Dial by your location

        +1 647 558 0588 Canada

        +1 778 907 2071 Canada

Meeting ID: 843 6841 6374

Find your local number: https://uvic.zoom.us/u/ktGtEdhgs 

Abstract

Sensitive information that cannot be shared presents an obstacle to gathering centralized datasets for machine learning tasks. Federated learning (FL) remedies this by leveraging data held across multiple clients without exchanging any private data. However, extra steps must be taken to ensure model privacy during the inference phase in order to eliminate the risk that model weights will be reverse-engineered to compromise training data privacy. Owners of a model, either jointly trained using FL or conventionally trained, may also wish to keep their model private in order to retain a monopoly on its use and commercialization. Such models may be intended to be used on sensitive data which also cannot be shared with the model owner. In this presentation I will detail cooperative activation functions (CAFs). CAFs enable inference on homomorphically encrypted data without requiring changes to model training and architecture. CAFs maintain the privacy of both the inference data and model weights. Motivating this work were vulnerabilities we identified in a comparison approach for private inference, which compromises the privacy of model weights and, therefore, training data. We evaluated CAFs on the same dataset to measure their effect on model accuracy and compute time. The results indicate that CAFs were able to deliver lossless or near-lossless performance.

Presented by: Optimizing UAV Trajectory for Maximum Sum Rate Using Proximal Policy Optimization

Date: Friday, August 16, 2024
Time: 12:00 pm
Place: Remote Via Zoom
Supervisor: Supervisor: Professor Hong-Chuan Yang

Speaker: Yawen Li

Title:  Optimizing UAV Trajectory for Maximum Sum Rate Using Proximal Policy Optimization

Supervisor:  Professor Hong-Chuan Yang

Date:  August 16, 2024

Time:  12:00pm

Location: Join Zoom Meeting https://uvic.zoom.us/j/85728294590

Abstract:  This seminar presents a study on the optimization of unmanned aerial vehicle (UAV) trajectory using advanced reinforcement learning (RL) algorithms, specifically Proximal Policy Optimization (PPO). The primary objective is to maximize the communication sum rate between the UAV and ground users by formulating it into a Markov Decision Process (MDP). The study introduces an innovative approach of action elimination to enhance the learning efficiency of RL agents by preventing them from selecting actions that do not contribute to the mission’s success. This method proved crucial in helping agents achieve higher rewards and reach their destinations on time, thereby avoiding unnecessary explorations. Additionally, the research explores the impact of different reward functions on the learning dynamics and performance of the RL agents. PPO shows a marked preference for cumulative rewards, reflecting its design to capitalize on long-term benefits. A significant portion of the research was dedicated to hyperparameter tuning within the PPO framework, where variables such as learning rates, clipping ratios, and buffer sizes were meticulously adjusted to refine the learning process. This tuning not only enhanced the performance of the PPO agent but also offered valuable insights into the sensitivity of RL algorithms to their operational parameters. However, the study acknowledges limitations, including the simplification of environmental factors and the two-dimensional trajectory optimization. Future work is suggested to integrate more complex environmental models and consider three-dimensional trajectory planning to address real-world applicability more effectively.

Presented by: Ahmed A. Abouelkhaire

Date: Friday, August 23, 2024
Time: 10:00 am
Place: Zoom, link below.

Join Zoom Meeting:

https://uvic.zoom.us/j/81897394054?pwd=4iDbQHzZOorBlp1xMreHbvXxsRR59V.1

Meeting ID: 818 9739 4054

Password: 078447

One tap mobile

+16475580588,,81897394054# Canada

+17789072071,,81897394054# Canada

Dial by your location

        +1 647 558 0588 Canada

        +1 778 907 2071 Canada

Meeting ID: 818 9739 4054

Abstract:

This seminar explores the application of image-based deep learning models for malware classification, leveraging a subset of the extensive MalNet-Image dataset, which includes around 87,000 binary images from a base of 1.2 million binary images based on Android APK files. The core contribution of this work lies in the innovative use of multiple components that, as far as we know, have not been used before to tackle the malware classification problem. Harnessing the power of deep neural networks (DNNs), which have demonstrated exceptional capabilities in various classification tasks, we aim to enhance the accuracy and efficiency of malware detection. These include Feature Pyramid Networks (FPN) to handle the file size scale issue when converting to images and the application of data augmentation techniques like Mix-up and Trivial Augment. We employ transfer learning with pre-trained models on ImageNet and optimize them using the AdamW Schedule-Free optimizer. Our experimental results show that the integration of these techniques achieves remarkable improvement in classification accuracy, with our best model achieving an F1 score of 0.6927 compared to 0.65 reported on the provided split for MalNet-Tiny. This could be considered a step forward in the field of malware classification using image-based deep learning models.