What should be tackled to prevent abuse of AI?

ByMany Wonderful Artists

Artificial intelligence (AI)andMachine learningTechnology has evolved at an unprecedented speed and has many useful applications ranging from machine translation to medical image analysis. Until now, artificial intelligence was used in malicious applications, but little attention was paid. there"OpenAI"The paper published by Miles Brundage et al. And others studied security risks when artificial intelligence technology was used with malicious intent and proposed a method to correctly predict these threats and to prevent or minimize damage It is.

[1802.07228] The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation
https://arxiv.org/abs/1802.07228

Preparing for Malicious Uses of AI
https://blog.openai.com/preparing-for-malicious-uses-of-ai/

If appropriate countermeasures against attacks using AI are not taken, it is necessary to first consider "what kind of attacks can be taken". Therefore, OpenAI recommends the following four countermeasures.
1. We will work closely with technical researchers to investigate, prevent and mitigate AI abuse.
2. If AI researchers and engineers have foreseen the possibility of becoming a harmful application due to abuse, contact with the person in charge at the relevant place in advance and discuss countermeasures.
3. Combine computer security technology and AI in case the AI ​​is exploited.
4. When discussing these issues, do not do it in a small community such as within a group or within a teamStakeholdersExpanding the scope of experts related to the scope of.

It is expected that the situation of the threat will change when the function of AI becomes stronger. Normally, if you are vulnerable to software or web services, careful preparation such as "Investigation of bug", "Preparation of attack code", "Confirmation of feasibility", etc. is required before actually going to attack I will. However, it becomes possible to drastically reduce the cost of these preparations by making advance preparations and attacks possible by AI, the appearance of "new kind of attacks" using AI and the "complicated attack method" It is said that there is a possibility that it may develop into a terrible situation that make it difficult to reach the real culprit, such as not being even noticed attacked by

ByKyle McDonald

OpenAI focuses on the following three areas as threats.

· Digital Security
Using AICyber ​​Attack, It becomes possible to make an appropriate attack according to the target, and it becomes possible to expand the attack target easily. In this way,Spear phishingThere is a possibility that the damage by the attack such as the attack may increase. Moreover, by spoofing the speech synthesis technology by AI, impersonating as a person in the company, such as to ask for a passwordSocial engineeringIt is also pointed out that the dangers of it will increase.

· Physical security
Attacks using thousands of unmanned aircraft and other physical systems and "crackingIt is possible to think that terrorist acts using AI may be carried out, such as placing a bomb on a robot cleaner that has done and exploding on the side of the VIP.

· Political security
AI operated surveillance cameras to collect indoor and outdoor data, analyzed to grasp the placement of security guards and the situation inside the building, and "to operate public opinion by embedding fake news in the broadcast It is said that there is a possibility to be able to launch an attack such as. They areAuthoritarianismAlthough it is considered to have a high impact on the nation, it points out that there is a danger that democratic nations will also lose their legitimate debate.

ByMike MacKenzie

Based on the above threats, we are proposing the following four issues to be addressed by AI engineers in the future.

· Cyber ​​Security
In the computer security community, we are implementing many measures related to AI. AI engineers need to apply know-how such as security techniques and "disclose vulnerabilities" before their research becomes a real threat. This means that it will lead to the prevention of the tragedy that makes research accomplished by struggle "enemy of mankind".

·Risk assessment
AI andMachine learningAs threats due to abuse increase, it is important to carry out risk assessment in advance.

·responsibility
AI researchers are obligated to improve the security level of AI itself. We also need to consider social frameworks such as education, ethical statements, standards, frameworks, norms.

Technical or policy solution
It seems necessary to examine room for investigation and promising technologies promising to make safer AI possible. In addition to protecting privacy and public security, we also consider countermeasures against regulations as items to consider.