JUNE 2023
|
|
“Training with malicious teachers: Poisoning attacks against machine learning” by Antonio Cinà (CISPA Helmholtz Center for Information Security) Date : 16 june 2023 Training cutting-edge machine learning (ML) models is often prohibitive to most because it requires expensive hardware and a huge amount of labeled data. To address this shortcoming, pre-trained models or publicly-available data are employed to reduce the financial development costs. However, these practices are becoming the Achilles’heel of the ML development chain because they expose the models to poisoning. These attacks assume the capacity of the attacker to tamper with the model training or data collection phases to drive the model toward unexpected misclassifications at test time. Because of the harm they can cause and the difficulties in detecting or mitigating them, poisoning attacks are nowadays considered the most feared threat by companies. Therefore, mindful monitoring of the data gathering and model training procedures is becoming imperative. In this talk we will categorize poisoning attacks according to their assumptions and attack strategies, showing that although poisoning has been acknowledged as a relevant threat in industry applications, acomplete systematization and critical review of the field was missing. We then inspect how poisoning can influence other ML aspects, going beyond misclassification violations and extending the attack surface in ML development. Finally, we shed light on the current limitations and open research questions in this field and propose possible future research directions. |
JUNE 2023
|
|
“Managing our online account security ” by Sasa Radomirovic (University of Surrey) Date : 23 june 2023 The number of online services, accounts, apps, and devices that we use is constantly increasing and so is the complexity of the interconnections between them. These interconnections have been exploited in attacks that range from account takeovers to cryptocurrency theft. Protecting users from such attacks is difficult because each user has a unique account ecosystem whose characteristics and variability we do not fully understand yet. In this talk I will introduce account access graphs which are a formal model to represent a user’s account ecosystem, i.e., the collection of accounts, apps, and devices, as well as their interconnections. I will show examples of real account access graphs and the first insights we have gained from them. I will then discuss some of the challenges we must overcome in order to build an account management tool that will empower users to better protect their account ecosystem. This talk is based on joint work published at CCS 2019, CHI 2022 and carried out at ETH Zurich, the University of Dundee and Heriot-Watt University. |
