5 mins: PaSSw0rdVib3s! - People share passwords, and even if someone accidentally sends you a password, you immediately know it is a password. Pa$$word123! does look more like a password than just the word password. 5 mins: The forensic context - why are we looking for passwords in devices? People reuse passwords all the time and even knowing how they typically construct their passwords can be useful in a criminal case (e.g. some people always put !s in their passwords). This information can help us break into other devices and/or accounts to gather evidence for the case (please note that this is only done in serious criminal cases). 6 mins: the data - positive examples are famous password leaks (note the drawbacks of these datasets). Negative examples: everything else you can find in a phone 2 mins: the model, I do want to mention it (after all it's a machine learning model that we built) but I won't elaborate too much. Honestly the real challenges of this project were in finding suitable data to train the models, and choosing the right way to evaluate them, so I want to focus on that. I think that makes the talk more suitable for anvaudience that is unfamiliar with machine learning too. 7 minutes: Evaluation - classification metrics versus ranking. How do you know which model is better? Is "Pa$$word123!" really more of a password than "password"? It definitely is more of a password than "#E29A86;"></div>" or "the". Finding a metric that reflected that was quite a challenge. I'll share our considerations. 5 mins: questions and discussion