How to Avoid Potential Security Risk by Securing Machine Learning Systems

August 4, 2020 Lenny Simon Lenny Simon

3 Min Read

If you are an organization that uses machine learning (ML), plans to use ML, or works with vendors that use ML, then read on. ML has been around for some time and has made significant progress on many tasks, including machine translation, autonomous vehicle control, image classification, video games, etc. However, until now, not much focus and attention has been on the security of ML systems themselves. We’re going to deep dive into the top 5 machine learning security risks and give you some insight on how to secure ML. But first, let’s talk about ML and some related terminology to lay a foundation.

We’re going to deep dive into the top 5 machine learning security risks and give you some insight on how to secure ML. But first, let’s talk about ML and some related terminology to lay a foundation.

Risks of Remote Work

Before we talk about what is Machine Learning (ML), let’s discuss some related terminology to clear things up because terminology in this space has been misused.

AI (Artificial Intelligence) – is a broad concept defined as a science of making things smart where machines perform humans tasks. The main point here is that AI isn’t precisely machine learning or smart things; it’s a concept where something can carry out human tasks.

ML (Machine Learning) – an approach (one of many) to AI where a system can learn from experience. Not only intended for AI goals (simulating human behavior) but can reduce the efforts or time humans spend on both simple and complex tasks. In simplest terms, ML is a system that can recognize patterns by using examples rather than by programming them. For example, if you have a system that consistently makes decisions and changes its behavior, then its ML.

DL (Deep Learning) – is more of a set of techniques for implementing ML that can identify patterns of patterns, like image recognition. These DL systems would identify primarily object edges, a structure, an object type, and then finally, an object itself.

Keep in mind that these definitions show that the cybersecurity field refers mainly to ML and not to AI, where a large number of ML tasks are not human-related.

Top 5 Security Risks

1. Data Poisoning

Since an ML system learns to do what it does directly from data, data can play an outsized role in the security of ML. If an attacker was successful at intentionally manipulating the data being used by an ML system, the entire ML system could be compromised. Since a data-poisoning attack requires special attention, ML engineers should understand what fraction of the training data an attacker can control and to what extent since several data sources are subject to poisoning. It’s helpful to understand the primary sources and consider how the data is stored, transferred, and processed. For example, raw data and data sets merged to test, train, and validate an ML system. This also should bring to light the fact of how sensitive data is to the workings of an ML system.

2. Online System Manipulation

Once an ML system runs correctly and continues to learn during production use and to modify its behavior over time, the ML system is said to be “online.” A smart attacker can push the ML system, which is still learning, in the wrong direction on purpose via system input. Doing so can slowly retrain the ML system to do the wrong thing and cause issues.

3. Adversarial Examples

In this type of risk, the idea is to fool the ML system by providing input involving small instances to make the ML system produce false predictions and categorizations.

4. Transfer-Learning Attack

One way to construct an ML system is through the act of tuning by using an already trained base model. A transfer attack can present a significant risk in this scenario as the attacker may devise an attack using those methods on your model. There is also the possibility that the ML system you are fine-tuning could be a Trojan horse that includes ML behavior that is anticipated.

5. Data Confidentiality

Protecting data is a challenge before adding the ML system to the mix. Now, with ML systems in place, there are subtle but effective data extraction attacks that can be an essential risk. Preserving data within an ML system is difficult because an ML system trained on confidential data will have certain aspects of those data built right into it through training. That said, attacks to extract that data from the ML system is common.

Securing Machine Learning

ML systems engineers can devise and field a more secure ML system by carefully considering the risks while designing, implementing, and fielding their specific ML system. Following these ten basic security principles can also help reduce risk to ML systems.

Principle 1: Secure the Weakest Link

Principle 2: Practice Defense in Depth

Principle 3: Fail Securely

Principle 4: Follow the Principle of Least Privilege

Principle 5: Compartmentalize

Principle 6: Keep It Simple

Principle 7: Promote Privacy

Principle 8: Remember That Hiding Secrets Is Hard

Principle 9: Be Reluctant to Trust

Principle 10: Use Your Community Resources

Learn More Today

Compuquip’s team of MSSP engineers can assist your organization with any cybersecurity concerns you might have. Want to talk to an expert on how you or your organization can secure your network with machine learning? Contact us today, and we’ll be happy to help!

how-to-choose-the-right-mssp-guide

cdo-guide-to-omnichannel-security