Data Poisoners Attack Automation

[rakhirhif8963]

But what happens when the learning process is automated? This doesn’t happen very often during development, but there are many cases where you want models to continually learn from new live data: “on-the-job” learning. At this point, it’s easy for someone to create “fake” data that feeds directly into AI systems and causes them to make erroneous predictions.

Consider, for example, the recommender systems of Amazon or Netflix. Recommendations can be easily changed by purchasing the product for someone else. Or bots can be created that rate programs or products millions of times. This will clearly change the ratings and “poison” the recommender system. Poisoning data is especially easy if those involved know that they are dealing with a self-learning system, such as a recommender system. All they need to do is make their attack “smart” enough to pass automated data checks, which is usually not very difficult.

Another problem is that data poisoning can be a long, slow process. Hackers can take their time changing data, injecting it several at a time. Moreover, this is often more effective, since a piecemeal change is dominican republic mobile database to detect than a one-time massive influx of data, and much harder to reverse.

How to Prevent Data Poisoning: Four Steps
To prevent poisoning, organizations can take the following steps:

Create an end-to-end ModelOps process and monitor all aspects of model performance and data drift using modern model management tools;
Create a business flow using workflow management tools to automatically retrain models. This means that before an updated version of the model can be used, it will need to go through a series of checks and validations performed by employees of business units;
Hire experienced data scientists and analysts. Many people mistakenly believe that software engineers can solve all technical issues, especially in the context of a shortage of qualified and experienced data scientists, but this is not true. We need experts who truly understand AI systems and ML algorithms and know what to look for when we deal with threats such as data poisoning;