AI Security Issues and Research
A series of security incidents have occurred in artificial intelligence applications such as autonomous driving and customer service robots, and have aroused people's concerns about the prospects of artificial intelligence applications. Chatgpt, which exploded last year, also had "thoughts of fleeing by itself" under the temptation of people. All these phenomena show the necessity of research on the safety of artificial intelligence
From the perspective of discipline development, there is a close relationship between artificial intelligence and cyberspace security, and artificial intelligence security is the inevitable result of the intersection of these two disciplines. Artificial intelligence theory and technology have effectively improved the intelligence level of cyberspace security attacks and defenses; more and more artificial intelligence models and algorithms have been found to have loopholes and security risks, which has led to artificial intelligence applications becoming a new era of cyberspace security. question
Artificial intelligence security is an interdisciplinary subject between artificial intelligence and cyberspace security. The two disciplines have already established profound theoretical and technical systems, and further seeing the intersection and logical connection of the two disciplines is the key to understanding and learning the safety of artificial intelligence
The concept of artificial intelligence has been developed with the development of computing technology since it was proposed in 1956. The scope of artificial intelligence theory and technology is constantly expanding. In terms of basic theory, it mainly includes knowledge representation, search, reasoning, intelligent computing, machine learning theory, representation learning, intelligent planning, neural network, etc. In terms of application fields, it mainly includes domain knowledge discovery and data mining, expert systems, natural language processing, image processing, intelligent voice, biometrics, computer vision, intelligent robots, etc.
Different schools of thought have different opinions on artificial intelligence research, thus forming their own doctrines. They think and study artificial intelligence theories and methods from different perspectives. Schools with greater influence and the doctrines they pursue,
Major: Semiotics (Symbolicism)
Bionic or Physiological School (Connectionism)
School of Cybernetics (Behaviorism, or Evolutionism)
Various schools of thought have proposed a large number of artificial intelligence theories and methods, but no matter which school of thought has paid little attention to the safety of artificial intelligence. Until recent years, driven by big data, various intelligent models have been established. The security issue of artificial intelligence has gradually attracted people's attention, which is determined by the processing of big data and the technical architecture of artificial intelligence itself.
AI driven by big data
Artificial intelligence models require data to be continuously updated, and the difficulty of updating and managing big data also greatly increases the security risk of data for artificial intelligence applications. Typical data security issues include: excessive data collection, data bias discrimination, abuse of data resources, data forgery, threats to fairness and justice, privacy leaks, and big data killing, etc.
The Hierarchy of AI Safety
Students who are familiar with computer networks are very familiar with the two models in the figure below, which is one of the reasons why artificial intelligence security and network security are closely related
Vulnerabilities in AI Security
There are vulnerabilities or security holes in information technology, which is the root cause of security problems. Artificial intelligence technology is no exception, and the security problems of artificial intelligence also lie in its own loopholes
Fragility of Data Distribution Assumptions
Vulnerabilities in the Model Update Mechanism
Openness in data processing
Computing Platform Vulnerabilities
Vulnerabilities in AI Decision Making
Essential Attributes of AI Safety
The basic attributes of information security include: confidentiality (Confidentiality), integrity (Integrity), availability (Availability), controllability (Controllability) and non-repudiation (Non-Repudiation)
As a kind of information system, intelligent system not only has the basic attributes of information security in the traditional sense, but also has unique attributes at the model algorithm level
Artificial intelligence security technology system
Among them, the two parts of artificial intelligence model and algorithm attack, artificial intelligence defense and governance are related to narrow artificial intelligence security. They focus on the attack methods and defense methods of various machine learning models such as supervised learning and unsupervised, and are a key part of the artificial intelligence security system.
In the adversarial attack and defense of the model, the main purpose of the attacker is to destroy the machine learning system, disrupt the detection and identification function of the machine learning system, and steal privacy from the machine learning system. As a security research, attack and defense are two important perspectives
From the perspective of attack, analyze the attacker's behavior, attack assumptions and goals, and study the techniques and methods to achieve these attack goals
From the perspective of the defender, study the corresponding defense methods for various attack behaviors
AI-safe data processing
The specific methods include small sample learning, unbalanced data processing and noise learning.
It has certain reference value for solving the main problems in network information security attack and defense, counter attack defense and artificial intelligence security governance.
Small samples, imbalance and noise problems are three typical AI data problems. The reasons for these three problems, on the one hand, are problems from the business data itself. For example, some types of sample data in intrusion detection are not easy to obtain, and users make label errors due to fatigue when labeling data.
On the other hand, the reason is from the attack behavior in the adversarial attack environment. The attacker adds noise to the training data, maliciously modifies the sample label, and selects a specific sample distribution during machine learning model training to realize the attack on the machine learning system. s attack
Artificial intelligence for attack and defense of network security
Network layer: Use the ports of hosts and network devices to perform network layer intrusion detection, denial of service attacks, etc.
The application layer involves more and more complex security issues, including security issues of various applications, as well as content security and behavioral security. Such as spam detection, fraudulent use of social network accounts, hot spots of public opinion, etc.
AI against attack and defense
Examples: financial fraud detection, intrusion detection, etc.
The core components in detection are usually various models based on artificial intelligence, so models and related data become important knowledge for attackers, and will inevitably become an important way and place for confrontation
The artificial intelligence counterattack technology system includes theories and attack methods, of which the theoretical part is mainly aimed at some basic models of artificial intelligence and related algorithms
From the perspective of attackers, many attack methods against artificial intelligence models have been discovered and classified from multiple perspectives such as attack purpose and attack knowledge.
Machine Learning Privacy Attack and Protection
When the training data contains personal sensitive attributes, privacy protection in the process of machine learning becomes an important issue for the success of machine learning
The privacy attack on machine learning is a special adversarial attack, which focuses on the sensitive acquisition of training data and the inference of model parameters, and its attack purpose is different from general adversarial attacks
The ways of machine learning privacy attacks are diverse, and for different machine learning systems, privacy attack methods have some similar methods. At the data level, there are also privacy attack methods for various types of data
From the perspective of machine learning privacy protection technology, it can be generally divided into two categories, namely, privacy protection from the perspective of data and privacy protection from the perspective of computing architecture.
Privacy Protection from a Data Perspective
The privacy protection of the data layer, which is the privacy protection for the original data, is also the starting point when the privacy protection method starts to be researched
The privacy protection of the model, the machine learning model is composed of the model structure type, model parameters and its operation rules. In comparison, the model structure type is easier to guess, so the model parameters become the sensitive information of the model
Privacy protection from the perspective of computing architecture
Machine Learning as a Service MLaaS: Deploy machine learning models in a cloud computing environment and provide users with calls in the form of services
In the distributed computing mode, how each unit participating in the calculation shares its own private data, different methods produce different privacy computing architectures
Secure multi-party computing, federated learning, etc.
AI Security Governance
From a technical point of view, the purpose of artificial intelligence governance is to prevent the unconscious misuse and conscious abuse of artificial intelligence technology in practical applications
data bias fairness justice data labeling quality artificial intelligence ethics
Artificial intelligence platform security
Driven by cloud computing technology, artificial intelligence is being presented as a service. Among them, the most typical is machine learning as a service
Network operation security issues The security issues of the artificial intelligence platform itself are mainly reflected in the monitoring and control of security risks in the calling and use of models and APIs in the platform. Data Security Issues
Mathematical basis
Probability Theory and Mathematical Statistics Linear Algebra Optimization Methods Discrete Mathematics