Blog Post
What Are the Unsafe Aspects of Machine Learning?
Security

What Are the Unsafe Aspects of Machine Learning?

Imagine asking your phone to suggest a movie for your next family night or allowing your car to navigate through the morning traffic to work. These scenarios, which are every day realities now, are all powered by artificial intelligence heralded by machine learning (ML).

Its resounding success and integration into our daily lives are undeniable. But just like any emerging technology, machine learning raises questions regarding safety and security. So, let’s pull back the tech curtain and delve into these concerns.

The Unseen Challenges Impacting Machine Learning Directly

Any ML model is only as good as its predictive ability. Ideally, the more iterations a model goes through, the better its predictive capabilities become. However, diverse problems can and often do arise that throw the model off. Some happen due to negligence or inherent biases. Others can be the result of malicious intent.

Unintentional risks

The quality of the data ML models train with is the most significant predictor of their usability. Unsurprisingly, “garbage in, garbage out” is one of the industry’s staple sayings. Basing the model on incomplete, incompatible, or compromised data puts any of its results into question.

Then there’s the problem of overfitting. It happens when models conform well to training data but do poorly in real-world scenarios with different parameters. For example, a stock market analysis tool that uses ML based on data collected during a period of calm and growth may not offer useful financial advice in times of crisis.

Even though ML developers try to obtain the most relevant and accurate training data, sometimes what they work with can become outdated before they can put their model to use.

Let’s say researchers develop a model that successfully predicts the mortality rates for people 75 and up. In the meantime, a medical breakthrough happens that helps prevent hypertension. Since training data and the new reality don’t align, the model isn’t irrelevant.

Biases are among the most widely debated machine learning risks. Some might be the result of researchers’ preconceptions. It’s more likely that the data itself results from systemic, cultural, and historical biases, though. There are already plenty of examples.

One of the most notorious is COMPASS. It is a tool used in several American states to help courts assess the likelihood of someone becoming a repeat offender. Based on data collected by the courts, COMPASS predicted recidivism rates for African Americans to be twice as high as those for Caucasian ones. That’s the “correct” conclusion based on what COMPASS had access to. Still, it doesn’t account for any context that might have contributed to the created data.

Adversarial attacks

Machine learning’s impact on the world comes with increased exposure of its vulnerabilities. Malicious actors have already devised several attack types that either tamper with or steal ML models to undermine the technology or for financial gain.

  1. Poisoning attacks: These introduce impurities into ML models’ training data. They’re challenging to pull off since the attacker needs access to the training data and ML algorithms. Such attacks are also effective since they can target and alter specific behaviors.
  2. Evasion attacks: These are easier to pull off and more common. They’re similar to poison attacks but happen after an ML tool’s deployment. Attackers who understand the tool’s classification principles can influence them. A famous example involves stop signs and a post-it note.

Researchers from NYU working on a road sign recognition algorithm were able to confuse it by placing a blank Post-it note on a stop sign. The algorithm had never encountered such an example before, so it misinterpreted the stop sign as a speed limit sign. While academic, this attempt shows how little outside influence it takes for ML to behave in potentially disastrous ways.

Associated Risks & Ways to Overcome Them

So far, we’ve only talked about machine learning without addressing associated concerns like cybersecurity. ML’s dependence on large-scale data collection means confidential or sensitive information is often on the line. Researchers and developers need to safeguard not just models, but the entire digital ecosystem they’re a part of.

Isolating their inputs, trials, and refinements from outside influences is paramount. The strongest defense would involve the development of closed-off networks. That’s impractical since team members still need to communicate, and third parties own most data sources accessible only online.

Machine learning development is an industry that benefits from tools like virtual private networks. They substitute the public and monitored default connection that internet service providers (ISPs) provide with one protected by layers of military-grade encryption. It’s essential because many types of malware could affect computers without a VPN. Since no one can log or observe what the developers do when using a VPN, no one can exploit their online activities to gain access to their research.

Encryption should also include storing source code, research results, and methodology documentation. This is especially relevant for startups, and other enterprises that need to balance exposure for funding and growth against the increased security and privacy needs that working with ML models entails.

Periodical Training for Employees

People engaged in ML need comprehensive training that covers detecting and avoiding the pitfalls of compromised or flawed models. Their cybersecurity education shouldn’t take a backseat either. Ignorant or negligent humans remain the part of any cybersecurity strategy that’s hardest to plan for.

Access to APIs, collaboration & knowledge tools, and countless other services for smooth development require accounts. One of the behaviors cybersecurity awareness should stomp out is using the same or weak passwords for them.

A password manager that automates and ensures reliable password implementation is a small expense. Together with multifactor authentication, it ensures data breaches that might cripple companies ML researchers are working with don’t affect their intellectual property.

Future Implications and Trends

Machine learning is rapidly evolving, creating new possibilities and challenges. As tech companies invest more resources into AI, we can expect advancements in countering ML’s safety risks. However, we must also remain vigilant against new risks emerging with such developments. We’re at the precipice of great leaps forward in machine learning — let’s ensure we continue to embrace improvement and preparedness.

Tech giants like Google, Apple, and Microsoft are investing billions to advance AI and machine learning. Notably, they are focusing on improving security measures in machine learning, reflecting industry-wide recognition of ML’s associated risks. For a peek into Google’s AI principles that guide their approach to AI development, including safety, check here.

Expert Opinions

Let’s call on the wisdom of industry leaders: Google’s Director of Research, Peter Norvig, once said, “Machine learning is the hot new thing.” While Jessica Hullman, a Northwestern University professor, cautioned, “Just because a machine learning model made a certain prediction, it doesn’t mean you should take it at face value.” These viewpoints provide an insightful exploration of machine learning’s potential and possible pitfalls.

Further Reading and Exploration

Machine learning is a rapidly evolving field, with potential risks being identified and tackled continually. Intrigued readers can delve deeper by checking out some of our related articles:

  1. Low-Code Security: An insightful overview of how low-code software reshapes the tech landscape and the associated security measures can be found here.
  2. A Guide to Data Security Compliance Laws and Regulations: Staying compliant with ever-changing data laws is challenging. Gain a comprehensive understanding at this link.
  3. Cybersecurity and Open-Source: For a deeper understanding of the pitfalls and promise of open-source software, our article here offers a comprehensive review.
  4. Securing Remote Access: As remote working becomes more prevalent, the question of securing remote access becomes more imminent, discussed here.
  5. SSL Certificates: Understanding SSL Certificates is fundamental in ensuring digital security. Learn more about the management of SSL certificates here.

For every tech enthusiast keen on maintaining security in this digital landscape, we encourage visiting our Security section for more knowledge.

Conclusion

Machine learning, a marvel of modern technology, has its fair share of challenges. However, understanding these pitfalls and making conscious moves toward safety can help us navigate this exciting terrain carefully. Machine learning’s potential excites us, but let’s not forget due diligence in its application.

  • Machine learning, underpinned by quality data, carries unintentional risks like overfitting, outdatedness, and bias.
  • Adversarial attacks, both poisoning and evasion types, threaten machine learning security.
  • Implementing integrated cybersecurity, network isolation, encryption, comprehensive training, and secure access can help counter these risks.

Remember, the safety of machine learning, like any potent tool, is in the hands of its users. Let’s use it responsibly.

Related posts

Leave a Reply

Required fields are marked *

Copyright © 2022 Blackdown.org. All rights reserved.