Next Seminar on 10.05.2023

Written on 04.05.2023 11:15 by Niklas Medinger

Dear All,

The next seminar(s) take place on 10.05.2023 at 15:00 (Session A) and 14:00 (Session B).

Session A: (15:00-15:30)
Syed Taqi Abbas Rizvi

Meeting-ID: 967 8620 5841
Kenncode: BT!u5=


Session B: (14:00-15:00)

Johannes Hägele, Shayari Bhattacharjee
(Temporary Link)


Session A:

15:00 - 15:30

Speaker: Syed Taqi Abbas Rizvi


No information provided.


Session B:

14:00 - 14:30

Speaker: Johannes Hägele
Type of talk: Bachelor Final
Advisor: Prof. Zeller
Title: Debugger Driven Input Grammar Mining on Embedded Systems
Research Area: RA5: Empirical and Behavioural Security
Automated security testing is indispensable for modern software development. Advances
in digitalization and the complexity of code bases increase potential attack vectors, so
that manual security testing or even auditing is impractical and insufficient. Fuzz testing
is a technique to automatically detect bugs in software. The first fuzzers fed random
inputs to the software under test but fail if the input structures are very complex or have
to pass a parsing stage. Generation-based fuzzing can overcome this barrier. Here input
specifications of the target program, are used to generate valid, or almost valid, random
test inputs.
However precise input specifications for programs are often outdated or even unavailable.
Automated generation of such precise input specifications is therefore a hot research topic.
In this work, we combine a set of sample inputs, software under test in binary format,
GDB the GNU Debugger, and the Mimid algorithm to automatically synthesize a human-
readable context-free grammar capturing the input language of the program under test.
The major benefit of our method is that it works in any environment with GDB access,
and therefore even on proprietary binaries in embedded systems or microcontrollers.


14:30 - 15:00

Speaker: Shayari Bhattacharjee

Type of talk: Master Final

Advisor: Dr.Nils Ole Tippenhauer, Prof. Martina Maggio

Title: Adversarial Robustness of Camera-Lidar based Multi-Sensor Fusion Architectures in Autonomous Driving

Research Area: RA4

Autonomous Vehicles(AV) have become a active research domain in the recent years. There have been significant contributions in the area of perception, planning and control using Machine Learning(ML) related to autonomous vehicles and recently, security concerns have also caught attention in the recent times owing to the development in the attacks on Machine Learning components. As a defense technique to counterfeit Adversarial ML attacks, Multi-Sensor Fusion(MSF) was proposed where inputs of multiple sensors present in the AV architecture are fused together to produce output with higher confidence. However, recently MSF scheme have also been shown vulnerable to certain attacks which is applicable to single or multiple sensor sources which causes system-wide effect and cause mispredictions or loss of accuracy.

In this thesis, first we conduct a study of various autonomous driving datasets and 2D/3D object detection frameworks. We also do a elaborate study of various attacks that have been used to attack image, point-cloud and multi-sensor fusion based detection frameworks. Furthermore, we test the various schemes of Camera-Lidar based Multisensor fusion based neural network against noise, corruptions and partial relay attacks.

In the course of the thesis, we developed and tested multiple versions of the attacks and evaluate them on the Camera-Lidar fusion frameworks in a black-box and white-box manner using KITTI validation dataset. This makes it possible to evaluate the success rate of the proposed attack on the various fusion schemes based on the fusion strategy, also evaluating the robustness of the fusion schemes against the proposed attacks.

Privacy Policy | Legal Notice
If you encounter technical problems, please contact the administrators.