US S&T’s Transportation Security Laboratory Evaluates Artificial Intelligence and Machine Learning Technologies
The Transportation Security Administration (TSA) screens 2.1 to 2.2 million travelers on an average weekday. Even though these numbers might be lower right now during the COVID-19 pandemic, the agency still needs to ensure that everyone traveling through nations airports can reach their destinations safely and securely. TSA and its employees need to be able to maintain effective and efficient security operations that can screen everyone for potential threats such as contraband, weapons and explosive devices. To help them better prepare for attempted security breaches, the Department of Homeland Security (DHS) Science and Technology Directorate’s (S&T) Transportation Security Laboratory (TSL) is evaluating artificial intelligence and machine learning (AI/ML) capabilities that have the potential to improve TSA’s ability to better protect air transportation system and everyone that relies on it.
A TSL employee demonstrates data collection methodologies that can be used to improve AI/ML technologies abilities to detect concealed threats on passengers and in their personal property.TSL is an S&T laboratory that performs research, development and validation of solutions designed to detect and mitigate the threat of improvised explosive devices. Recently, its subject matter experts (SMEs) have turned their attention to AI/ML technologies and algorithms for their potential to improve detection of concealed threats on passengers and their personal property. TSL’s current priority is to provide a valid certification test for these technologies and algorithms to vendors and also help vendors gather training data in the event that they fail this test. TSL is exploring AI/ML to see how to best advance their mission, while specifically focusing on the development of new protocols and tools that will effectively evaluate these new, embedded algorithms. TSL plans to collaborate with other DHS agencies to develop system requirements for diagnostics and visibility that will enable detection trustworthiness.
“Current security algorithms rely on manual data mining to identify complex threats and translate them into coded rules for existing screening technologies. However, machine learning is an emergent technology that has the potential to change how new security screening algorithms will be created,” said Barry Smith, Manager of TSL’s Applied Research Division. “These new algorithms will have the potential to outperform the existing rule-based algorithms and present significant opportunities for improving DHS missions that involve efficiently detecting threats that may be concealed among day-to-day travelers and commerce.”
SMEs in the air transportation security space recognize the potential that AI/ML can have in their field. Reduced first stage alarms in checked-bag screening systems; automatic screening system detection of weapons and other concealed threats in carry-on bags and on passengers; and reduced secondary passenger screenings and delays are among the most significant benefits that can be provided when algorithms are effectively used to improve existing air passenger screening technologies.
“Machine learning, while relatively easy to engineer, is quite difficult to train and test with our current protocols and tools,” said Lee Spanier, TSL Spectroscopy Branch Manager. “Machine learning-crafted algorithms possess unusual vulnerabilities and susceptibilities that are hard to discover without very large data sets and new methods.”
In order to evaluate AI/ML’s potential efficacy in the field, TSL is developing new tools, methods, and procedures to test and train these algorithms effectively and efficiently before they are commercialized. TSL is currently planning to implement the following actions:
- Create and implement a roadmap to compile plans and initiatives into a comprehensive strategy for ML training, readiness and detection testing;
- Provide educational and training opportunities for federal staff through collaborative efforts with Rowan University, the National Institute of Standards and Technology and TSA;
- Hire scientists and engineers with specialized skills in ML, data science and data reconstruction;
- Improve current ML training and testing protocols by developing and using synthetic training data sets and simulants that mimic the realistic and wide range of threats that could be concealed in baggage or in a passenger’s possession;
- Investigate techniques for using ML to validate threat simulants; and
- Develop new laboratory infrastructure to rapidly and comprehensively study, recreate, model, and understand the complex properties of threats to create synthetic augmentation libraries to fully train and test ML-crafted algorithms.
“Our AI/ML evaluation efforts will ensure that we understand what’s needed to make ML algorithms work effectively,” said Smith. “Figuring out their weaknesses and eliminating them is how we can validate these algorithms and be sure that they will correctly detect relevant threats in our air transportation system. Ultimately, our goal is to solve the training and quality assurance challenges so that the performance gains of ML-crafted algorithms can be realized, drastically improving our existing screening technologies and the security of commercial aviation.”
TSL is currently developing the aforementioned Machine Learning Roadmap to determine the capabilities, sequencing of activities, and resource requirements that need to be developed or improved as part of its efforts to evaluate AI/ML. They have also formed a working group to collaborate with TSA and others within S&T pursuing ML efforts. The progress made here has the potential to impact many homeland security missions.
“AI/ML algorithms have implications that reach beyond TSA,” said Spanier. “Other agencies, for example, Customs and Border Protection, are hoping to use these algorithms in their efforts to detect shipments of illegal drugs and contraband that are being smuggled into the country. Therefore, we need to make sure we thoroughly evaluate these algorithms so that other homeland security agencies can benefit from our efforts and effectively incorporate them into their existing safety protocols.”