ScamBuster
Identify Cybersecurity Threats - in real time, on your device
Created on 27th September 2024
•
ScamBuster
Identify Cybersecurity Threats - in real time, on your device
Describe your project
Two segments of the scope are:
- The project scope starts with the incoming text sms.
The scope ends with triggering a notification to the user mentioning that the text sms received is probably a scam. - In second scenario: the project scope starts with 'user clicking a CTA in the notification bar for taking the screenshot'
The scope ends with triggering a notification to the user to tell whether the screenshot depicts any kind of possible scam.
Kindly note - the identification is for a possible scam (and not a spam).
Scope of project starts once the App is installed on the user device. Scope covers:
- Accessing the content of incoming text message
- Accessing the content of screenshot that is triggered by the user via a CTA in the notification
- Analyzing the content (along with urls, if present) remotely via api.
- Returning a probability score of the likelihood of it being a "scam".
- Showing Notification on user device about it being a "scam" and hence, alerting the user.
Anything not covered under above is out of scope for this solution.
Future Opportunities:
- Extension of the scope to cover inbound calls received from a unknown number. The model will analyze the chain of conversation after every sentence and will continuously return the probability of likelihood of "scam".
- Real time analytics of voice for identifying and alerting for an AI generated voice call.
- Enhancement of training content for covering more and more cybersecurity threats.
- Anonymization of the data before it is sent to the api
- Read the incoming msgs on WhatsApp / Instagram via seeking access to the notification section of device
Challenges we ran into
We started the project with a vision to identify scams that are happening over voice calls such as, Fedex scam, Family member in trouble scam etc. But the major challenge we ran in to was not been able to access the real time call audio / transcription as operating systems do not allow access for those. We have the entire code ready and we tested it over the AI generated voice calls on local.
We, then, switched the current scope to doing this analytics for text / images and keep the voice in the future scope of the project.
Another challenge that we got in to was - when we read the project has to be executed via Gemini AI only. Somehow OpenAI (4o) was giving much better response than Gemini AI. We had to compromise on the accuracy while shifting to Gemini AI. We are sure that we will be able to improve the accuracy in future via extensive training of the model.
Tracks Applied (1)
17. Problem statement shared by Central Cyber Security Agency
Technologies used