Major security, privacy and ethical blindspots in AI development

Security, privacy and ethics are low-priority issues for developers when modelling their machine learning (ML) solutions, new research has found.

O'Reilly's 2019 AI Adoption in the Enterprise survey found security is the most serious blind spot within an organisation.

Nearly three-quarters (73%) of respondents indicated they don't check for security vulnerabilities during model building. More than half (59%) of organisations also do not consider fairness, bias or ethical issues during ML development. Privacy is similarly neglected, with only 35% checking for issues during model building and deployment.

Instead, the majority of developmental resources are focused on ensuring artificial intelligence (AI) projects are accurate and successful, the report found.

"AI maturity and usage has grown exponentially in the last year. However, considerable hurdles remain that keep it from reaching critical mass," says Ben Lorica, chief data scientist, O'Reilly.

"As AI and ML become increasingly automated, it's paramount organisations invest the necessary time and resources to get security and ethics right," he explains.

"To do this, enterprises need the right talent and the best data. Closing the skills gap and taking another look at data quality should be their top priorities in the coming year."

The majority (55%) of developers mitigate against unexpected outcomes or predictions, but this still leaves a large number who don't. Furthermore, 16% of respondents do not check for any risks at all during development.

This lack of due diligence is likely due to numerous internal challenges and factors, but the greatest roadblock hindering progress is cultural resistance, as indicated by 23% of respondents, O'Reilly says.

The research also shows 19% of organisations struggle to adopt AI due to a lack of data and data quality issues, as well as the absence of necessary skills for development. The most chronic skills shortages by far were centred around ML modelling and data science (57%). To make progress in the areas of security, privacy and ethics, organisations urgently need to address these talent shortages.

Other key findings include:

• The overwhelming majority of organisations (81%) have started down the route of AI adoption. Most are in the evaluation or proof of concept stage (54%), while 27% have revenue-bearing AI projects in production. 
• A significant minority (19%) of companies have not started any AI projects.
• Machine learning has emerged as the most popular form of AI used by enterprises. Nearly two-thirds (63%) use supervised learning solutions while 55 per cent are using deep learning technology. Model-based methods are used by almost half (48%) of respondents.
• AI is most likely to be used in research and development (R-D) departments (50 per cent), customer service (34%) and IT (33 per cent). Legal functions have seen the least innovation, with only 5% making use of AI technologies.
• TensorFlow (55%) and scikit-learn (48%) are the most popular AI tools in use today.