10:00am - 11:00am
Presenter: Mohammad Shadab
Advisor: Professor Dmitry Ponomarev
When: Tuesday, April 30, 2024 10 am
Where: PO3
Title: Timing Attacks in GPU Systems
Abstract: GPUs enhance the performance of computationally intensive applications. As GPUs are increasingly used in deep learning workloads, a number of vulnerabilities have emerged. These vulnerabilities can be exploited to extract information from a victim application by employing side channels and covert channels through shared resources. In this presentation, we will discuss hardware security in GPUs, particularly their vulnerabilities to timing attacks. There are several prerequisites to make attacks possible on GPUs, such as co-location and concurrency between two applications. To establish attacks on GPUs, we need some information which is not published by manufacturers. Therefore, one of the critical parts of designing attacks on GPUs is reverse engineering to obtain necessary information, such as scheduling policies or co-location algorithms. After conducting several practical reverse engineering on GPUs, the next step is extracting information from a victim model such as DNN model characteristics. In this presentation, we will discuss related works that establish covert and side channels and use them to design attacks on GPUs. We will also introduce a new attack surface on the GPUs to demonstrate a possible channel to communicate between two malicious applications or extract information from a victim model.
Advisor: Professor Dmitry Ponomarev
When: Tuesday, April 30, 2024 10 am
Where: PO3
Title: Timing Attacks in GPU Systems
Abstract: GPUs enhance the performance of computationally intensive applications. As GPUs are increasingly used in deep learning workloads, a number of vulnerabilities have emerged. These vulnerabilities can be exploited to extract information from a victim application by employing side channels and covert channels through shared resources. In this presentation, we will discuss hardware security in GPUs, particularly their vulnerabilities to timing attacks. There are several prerequisites to make attacks possible on GPUs, such as co-location and concurrency between two applications. To establish attacks on GPUs, we need some information which is not published by manufacturers. Therefore, one of the critical parts of designing attacks on GPUs is reverse engineering to obtain necessary information, such as scheduling policies or co-location algorithms. After conducting several practical reverse engineering on GPUs, the next step is extracting information from a victim model such as DNN model characteristics. In this presentation, we will discuss related works that establish covert and side channels and use them to design attacks on GPUs. We will also introduce a new attack surface on the GPUs to demonstrate a possible channel to communicate between two malicious applications or extract information from a victim model.
9:30am - 10:15am
EB T1 and Zoom. Link below.
2:00pm - 4:00pm
Presenter: Xiaohan Zhang
Advisor: Shiqi Zhang
Date: Thursday, May 2, 2-4 pm
Location: EB T1
Zoom link: https://binghamton.zoom.us/j/9953067186?omn=97023042413
Title: Symbol Grounding for Task and Motion Planning in Robotics
Abstract: For modern robots that are equipped with a set of skills, such as manipulation and navigation, it's crucial that they can autonomously make decisions to complete tasks over extended periods of time. To this end, researchers have been developing planning algorithms that allow robots to effectively sequence and use their skills to reach specific task-level goals. Existing planning systems have a strong reliance on structured and predictive world models which are often represented as symbol tokens. Symbol grounding is thus required to establish a meaningful connection between these abstract symbols and their real-world interpretations that robots can understand. A fundamental approach to symbol grounding for robots involves utilizing their perception capabilities, allowing the mapping of sensory readings to symbol tokens and thereby assigning perceptual meanings to symbols. Additionally, robots frequently need to plan their actions in the process of grounding, given robot actions being tightly coupled with perception. Motivated by the above observations, this dissertation focuses on 1) symbol grounding using perception and action, and 2) robot planning with grounded symbols. Our research draws on theories and methods from computer vision and machine learning with both being integrated with AI planning and continuous motion planning towards building long-horizon robot autonomy.
Advisor: Shiqi Zhang
Date: Thursday, May 2, 2-4 pm
Location: EB T1
Zoom link: https://binghamton.zoom.us/j/9953067186?omn=97023042413
Title: Symbol Grounding for Task and Motion Planning in Robotics
Abstract: For modern robots that are equipped with a set of skills, such as manipulation and navigation, it's crucial that they can autonomously make decisions to complete tasks over extended periods of time. To this end, researchers have been developing planning algorithms that allow robots to effectively sequence and use their skills to reach specific task-level goals. Existing planning systems have a strong reliance on structured and predictive world models which are often represented as symbol tokens. Symbol grounding is thus required to establish a meaningful connection between these abstract symbols and their real-world interpretations that robots can understand. A fundamental approach to symbol grounding for robots involves utilizing their perception capabilities, allowing the mapping of sensory readings to symbol tokens and thereby assigning perceptual meanings to symbols. Additionally, robots frequently need to plan their actions in the process of grounding, given robot actions being tightly coupled with perception. Motivated by the above observations, this dissertation focuses on 1) symbol grounding using perception and action, and 2) robot planning with grounded symbols. Our research draws on theories and methods from computer vision and machine learning with both being integrated with AI planning and continuous motion planning towards building long-horizon robot autonomy.
3:00pm - 4:00pm
Presenter: Huaxin Tang
Advisor: Professor David Liu
When: Thursday, May 2, 4 pm
Where: EB P03
Title: A Compiler Framework for Proactive UAV Regulation Enforcement
Abstract: In the rapidly evolving landscape of Unmanned Aerial Vehicles (UAVs), regulation enforcement is critical. Unfortunately, existing practices are largely manual and reactive in nature. We present THEMIS, a novel compiler-directed approach for automated and proactive regulation enforcement. By expressing regulations through a specification language and integrating their enforcement into the compilation process, THEMIS enables safe and regulation-compliant UAV flights by enforcing prohibited and restricted areas, avoiding flights over humans, and managing maximum limits of altitude and speed. Our framework features a bidirectional interface that allows the concrete algorithms used for enforcement to be customized. Our evaluation shows THEMIS-compiled autopilots can adhere to regulatory constraints amidst complex flight conditions, while significantly reducing the burden of UAV operators.
Advisor: Professor David Liu
When: Thursday, May 2, 4 pm
Where: EB P03
Title: A Compiler Framework for Proactive UAV Regulation Enforcement
Abstract: In the rapidly evolving landscape of Unmanned Aerial Vehicles (UAVs), regulation enforcement is critical. Unfortunately, existing practices are largely manual and reactive in nature. We present THEMIS, a novel compiler-directed approach for automated and proactive regulation enforcement. By expressing regulations through a specification language and integrating their enforcement into the compilation process, THEMIS enables safe and regulation-compliant UAV flights by enforcing prohibited and restricted areas, avoiding flights over humans, and managing maximum limits of altitude and speed. Our framework features a bidirectional interface that allows the concrete algorithms used for enforcement to be customized. Our evaluation shows THEMIS-compiled autopilots can adhere to regulatory constraints amidst complex flight conditions, while significantly reducing the burden of UAV operators.
4:00pm - 5:00pm
Presenter: Ryuma Nakahata
Advisor: Professor Kenneth Chiu
When: Thursday, May 2, 4 pm
Where: EB N03
Title: PtychoFormer: A Transformer-Based Deep Learning Model For Ptychography Phase Retrieval
Abstract: Ptychography is a computational microscopy technique that retrieves a sample's phase and amplitude contrast image from diffraction patterns resulting from coherent light beam illumination. The phase problem in ptychography is the challenge of recovering the phase information of a sample from intensity-only measurements of the diffracted waves. To address this problem, the collected diffraction patterns maintain significant spatial overlaps between illuminated areas in adjacent scan points. The redundant information from the overlaps aids in reconstructing the amplitude and the lost phase of the object. Traditional phase retrieval algorithms project successive diffraction patterns to recover high-quality images. However, their iterative nature hinders their ability to achieve real-time imaging. Our research introduces an instantaneous ptychographic phase retrieval algorithm called PtychoFormer. This image-to-image transformer-based model leverages a hierarchical multi-head self-attention to process sets of diffraction patterns to perform subregional reconstructions. Our findings show that the proposed model can effectively synthesize high-quality estimates with minimal spatial overlaps between adjacent diffraction patterns. We also demonstrate that PtychoFormer is a practical tool for ptychography, functioning as a real-time phase retrieval algorithm that exhibits competitive performance against a traditional ptychographic algorithm and outperforms prior deep learning methods.
Advisor: Professor Kenneth Chiu
When: Thursday, May 2, 4 pm
Where: EB N03
Title: PtychoFormer: A Transformer-Based Deep Learning Model For Ptychography Phase Retrieval
Abstract: Ptychography is a computational microscopy technique that retrieves a sample's phase and amplitude contrast image from diffraction patterns resulting from coherent light beam illumination. The phase problem in ptychography is the challenge of recovering the phase information of a sample from intensity-only measurements of the diffracted waves. To address this problem, the collected diffraction patterns maintain significant spatial overlaps between illuminated areas in adjacent scan points. The redundant information from the overlaps aids in reconstructing the amplitude and the lost phase of the object. Traditional phase retrieval algorithms project successive diffraction patterns to recover high-quality images. However, their iterative nature hinders their ability to achieve real-time imaging. Our research introduces an instantaneous ptychographic phase retrieval algorithm called PtychoFormer. This image-to-image transformer-based model leverages a hierarchical multi-head self-attention to process sets of diffraction patterns to perform subregional reconstructions. Our findings show that the proposed model can effectively synthesize high-quality estimates with minimal spatial overlaps between adjacent diffraction patterns. We also demonstrate that PtychoFormer is a practical tool for ptychography, functioning as a real-time phase retrieval algorithm that exhibits competitive performance against a traditional ptychographic algorithm and outperforms prior deep learning methods.
10:30am - 4:00pm
UU 111
Agenda to follow as the date draws nearer.
1:00pm - 2:00pm
Presenter: Williams Zhang Cen
Advisor: Professor Dmitry Ponomarev
When: Wednesday, May 8, 2024, 1 pm
Where: EB PO3
Title: Program Verification Using Side-Channels
Abstract: In this talk, we will discuss recent research related to ensuring the integrity of program execution to protect against a range of attacks. Specifically, we will focus on control flow integrity solutions, both software-only and with support of hardware. After that, we will review recent literature that verifies program execution remotely using analog side-channel signatures produced as a result of execution. Finally, we will articulate some initial ideas on how to use microarchitectural side-channels to remotely verify correct execution of programs, effectively putting side-channels to good use.
Advisor: Professor Dmitry Ponomarev
When: Wednesday, May 8, 2024, 1 pm
Where: EB PO3
Title: Program Verification Using Side-Channels
Abstract: In this talk, we will discuss recent research related to ensuring the integrity of program execution to protect against a range of attacks. Specifically, we will focus on control flow integrity solutions, both software-only and with support of hardware. After that, we will review recent literature that verifies program execution remotely using analog side-channel signatures produced as a result of execution. Finally, we will articulate some initial ideas on how to use microarchitectural side-channels to remotely verify correct execution of programs, effectively putting side-channels to good use.
10:30am - 1:30pm
University Union, Old Union Hall
10:30am - 11:30am
