Attack Model for Generic Intelligent Systems

Figures & data

Figure 1. Steps toward “intelligence.”

Figure 2. Basic architecture of intelligent systems.

Figure 3. Locations and types of data in the intelligent system architecture.

Figure 4. Machine learning process.

Table 1. Potential attacks on static data in intelligent systems.

Potential Attack	Potential Impact
Data theft	Data breach leading to unauthorized data access.
Data modification	Modified data can be incorrect, inaccurate, or damaged; thus, the intelligent system will not function correctly.
Data deletion	The data no longer exist; thus, the data cannot be processed by the intelligent system.

Table 2. Potential attacks on data in transit of intelligent systems.

Potential Attack	Potential Impact
Eavesdropping or data theft	Loss of data privacy and secrecy.
Data interception	Loss of data privacy and secrecy.
Data modification	Data may be incorrect, inaccurate, or damaged; thus, the intelligent system cannot function correctly.
Data deletion	The data no longer exist; thus, the data cannot be processed by the intelligent system.
Data injection	Unauthorized creation and transmission of data results in data incorrectness; thus, the intelligent system cannot function as expected.

Table 3. Attack methods and attack types on data in intelligent systems.

Attack Method	Attack Type	Explanation
Software vulnerabilities	Data theft	Software vulnerabilities can be exploited by an adversary to access, steal, and/or alter data.
	Data modification
Poor or no encryption	Data theft	Unencrypted data stored on a hard disk or in a database can be accessed easily.
	Data modification
	Eavesdropping
	Data interception
System misconfiguration	Data theft	Ineffective system configuration creates potential vulnerabilities for an adversary to find a backdoor or other entry point into the intelligent system.
	Data modification
	Eavesdropping
	Data interception
	Data deletion
	Data injection
Man-in-the-middle	Data theft	Man-in-the-middle attacks typically target data in transit. Here, data can be damaged or prevented from reaching the intended destination.
	Data modification
	Eavesdropping
	Data interception
	Data deletion
	Data injection
Denial of service	Data modification	When data are modified or deleted, the intelligent system cannot function properly. In addition, if data in transit are intercepted, the data will not reach the intended destination; thus, the intelligent system will not function as expected.
	Data interception
	Data deletion

Figure 5. Pairings between attack vector and data types.

Table 4. Attack methods on processes and components of intelligent systems.

Components	Attack Method	Example/Impact
Input data	Data poisoning	Machine learning and intelligent systems’ effectiveness is reduced. The system cannot function according to requirements.
	Untrusted data	When the data input to the system are not trusted and/or incorrect, the effectiveness of the machine learning process is reduced.
	Sensitive data	There is an increased risk of data being compromised and breached.
	Data modification	When the data input to the system are modified in expected ways, the effectiveness of the machine learning process is reduced.
	Insecure connection	Adversaries can access data transmitted over insecure channels.
Machine learning process	Model theft	If the learning model of an intelligent system is stolen, the secrecy of the algorithm will be exposed.
	Model modification	The effectiveness of machine learning and the intelligent system is reduced. The system cannot function according to requirements.
	Model deletion	The machine learning model or learning process become inaccessible, which means the intelligent system is unusable.
Intelligent system usage	Unauthorized access	When the number of unauthorized accesses is high, accessibility can be compromised or become nonexistent.
		Access to parts of the intelligent system by unauthorized entities produces the same effects described above for input data and the machine learning process.

Figure 6. Attack model of an intelligent system.

Table 5. Attack methods on data and security characteristics.

Attack Method	Affected Security Characteristics
Eavesdropping	Confidentiality
Data theft	Confidentiality
Data modification	Integrity
Data deletion	Availability
Data interception	Confidentiality
	Availability
Data injection	Integrity
	Availability

Table 6. Attack methods on the processes and components and security characteristics.

Attack Method	Affected Security Characteristics
Data poisoning	Integrity
Untrusted data	Integrity
Data modification	Integrity
Insecure connection	Confidentiality
	Integrity
Machine learning model theft	Confidentiality
Machine learning model modification	Integrity
Machine learning model deletion	Availability
Unauthorized access	Confidentiality
	Integrity
	Availability