Artificial intelligence must be wary of data traps

  The soldiers stand by fraud.

In the intelligent age, disruptive technologies continue to emerge, and new methods and forms of war deception continue to emerge.

"If you have your data, I can create various ways to deceive your artificial intelligence system." Research experiments show that once one party obtains the opponent's artificial intelligence training data set in an intelligent war, it can find its weaknesses and blind spots and To implement deception, artificial intelligence must be alert to data traps.

  At present, artificial intelligence analyzes and processes data much faster than human analysts, and can find behavior patterns and laws that are difficult for the human brain to find, but it can also make mistakes that the human brain does not make.

The reason is that machine learning algorithms must rely on large amounts of data for training. Data is to artificial intelligence what blood is to humans. Sharing data is more difficult than designing algorithms.

If the data set is too small, the data is inaccurate, or is maliciously tampered with by the opponent, then the effect of machine learning will be greatly reduced, and even be misled and misjudged.

Especially in the national security and military fields, harmful data can cause serious consequences.

Once the artificial intelligence training data set is mastered by the opponent, the opponent will design data traps, implement deception, provide fake data and induce artificial intelligence to learn wrong data.

What’s more serious is that because the internal working mechanism of machine learning algorithms is obscure and difficult to understand, people usually don’t know why artificial intelligence makes mistakes. Especially when there are no catastrophic consequences, it is difficult to detect artificial intelligence errors. Falling into a data trap at a loss.

  So, how to avoid data traps?

First, human brain intervention is required.

Only humans have the ability to label data classification, so they cannot simply throw the data to the machine algorithm, and hope that artificial intelligence solves all problems without human brain intervention.

If only a large amount of data is provided and there is no "smart human brain" that can distinguish the data, then artificial intelligence can only provide mechanical answers, not the correct answers people need.

Human brain intervention can not only ensure that the artificial intelligence obtains the correct data, it can also check whether it is learning the correct data.

Second, build a cross-domain team.

The "smart human brain" that can avoid data traps must come from a cross-field team, and computer experts, programmers, big data experts, and artificial intelligence experts must work closely with experienced professionals in related fields.

In the future, after the continuous development and maturity of artificial intelligence, it will be possible to directly provide real-time intelligence to warfighters. This requires warfighters to continuously provide feedback to the "smart brain" team to update and correct data in a timely manner.

Once again, cross-check multi-source data.

It is easy to be deceived by opponents to use one kind of sensor to detect a target. Therefore, it is necessary to use multiple sensors such as vision, radar, and infrared to detect the same target, and compare and verify data from different sources in order to distinguish the authenticity and discover hidden scams.

Furthermore, label the data classification.

At present, even advanced artificial intelligence will make absurd low-level mistakes and even mistake the toothbrush for a baseball stick.

Therefore, it is not possible to provide unprocessed raw data for machine learning, especially in the early stages of training. The machine algorithm should be provided with correctly classified and labeled real data to verify whether the conclusions of artificial intelligence are correct and ensure that artificial intelligence assists Decisions are accurate and efficient.

Finally, take adversarial learning.

Establish an intelligent blue army, develop artificial intelligence opponents, let the artificial intelligences that are opponents and confront each other fight each other, conduct adversarial learning in the process of wits, improve the ability to identify data traps in adversarial learning, and achieve victory by wisdom.

In short, current artificial intelligence is still inseparable from human brain control, and avoiding data traps ultimately depends on human experience and wisdom.

  Liu Peng