![](https://csdnimg.cn/release/download_crawler_static/15086700/bg4.jpg)
4
TABLE 2: A summary of attacks against server-based FL.
Attack Type
Attack Target Attacker Role FL Scenario Attack Complexity
Model Training Data Participant Server H2B H2C Attack Iteration Auxiliary Knowledge Required
One Round Multiple Rounds
Data Poisoning YES NO YES NO YES YES YES YES YES
Model Poisoning YES NO YES NO YES NO YES YES YES
Infer Class Representatives NO YES YES YES YES NO NO YES YES
Infer Membership NO YES YES YES YES NO NO YES YES
Infer Properties NO YES YES YES YES NO NO YES YES
Infer Training Inputs and Labels NO YES NO YES YES NO YES YES NO
2.3 Training Phase v.s. Inference Phase
Training Phase. Attacks conducted during the training
phase attempt to learn, influence, or corrupt the FL model it-
self [51]. During the training phase, the attacker can run data
poisoning attacks to compromise the integrity of the training
dataset [26], [28], [52], [53], [54], or model poisoning attacks
to compromise the integrity of the learning process [27], [55].
The attacker can also launch a range of inference attacks
on an individual participant’s update or on the aggregated
update from all participants.
Inference Phase. Attacks conducted during the inference
phase are called evasion/exploratory attacks [56]. They
generally do not alter the targeted model, but instead, either
trick it to produce wrong predictions (targeted/untargeted)
or collect evidence about the model characteristics. The
effectiveness of such attacks is largely determined by the
information that is available to the adversary about the
model. Inference phase attacks can be classified into white-
box attacks (i.e. with full access to the FL model) and black-
box attacks (i.e. only able to query the FL model). In FL, the
global model maintained by the server suffers from the same
evasion attacks as in the conventional ML setting when the
target model is deployed as a service. Moreover, the model
broadcast step in FL makes the global model a white-box to
any malicious participant. Thus, FL requires extra efforts to
defend against white-box evasion attacks.
3 POISONING ATTACKS
Depending on the attacker’s objective, poisoning attacks can
be classified into two categories: a) untargeted poisoning
attacks [55], [57], [58], [59], [60] and b) targeted poisoning
attacks [26], [27], [61], [62], [63], [64], [65], [66].
Untargeted poisoning attacks aim to compromise the
integrity of the target model, which may eventually result
in a denial-of-service attack. In targeted poisoning attacks,
the learnt model outputs the target label specified by the
adversary for particular testing examples, e.g., predicting
spams as non-spams, and predicting attacker-desired labels
for testing examples with a particular Trojan trigger (back-
door/trojan attacks). However, the testing error for other
testing examples is unaffected. Generally, targeted attacks
is more difficult to conduct than untargeted attacks as the
attacker has a specific goal to achieve. Poisoning attacks
during the training phase can be performed on the data or
on the model. Fig. 2 shows that the poisoned updates can
be sourced from two poisoning attacks: (1) data poisoning
attack during local data collection; and (2) model poisoning
attack during local model training process. At a high level,
both poisoning attacks attempt to modify the behavior of
the target model in some undesirable way.
Fig. 2: Data v.s. model poisoning attacks in FL.
3.1 Data Poisoning
Data poisoning attacks largely fall in two categories: 1)
clean-label [44], [65], [67], [68], [69], [70] and 2) dirty-
label [63], [64]. Clean-label attacks assume that the adver-
sary cannot change the label of any training data as there
is a process by which data are certified as belonging to
the correct class and the poisoning of data samples has to
be imperceptible. In contrast, in dirty-label poisoning, the
adversary can introduce a number of data samples it wishes
to be misclassified by the model with the desired target label
into the training data.
One common example of dirty-label poisoning attack
is the label-flipping attack [28], [58]. The labels of honest
training examples of one class are flipped to another class
while the features of the data are kept unchanged. For
example, the malicious participants in the system can poison
their dataset by flipping all 1s into 7s. A successful attack
produces a model that is unable to correctly classify 1s
and incorrectly predicts them to be 7s. Another weak but
realistic attack scenario is backdoor poisoning [24], [25],
[64]. Here, an adversary can modify individual features or
small regions of the original training dataset to implant a
backdoor trigger into the model. The model will behave
normally on clean data, yet will constantly predict a target
class whenever the trigger (e.g., a stamp on an image)
appears. In this way, the attacks are harder to be detected.
For backdoor attacks in FL, a global trigger pattern can be
decomposed into separate local patterns and embed them
into the training set of different adversarial participants
respectively [25].
Data poisoning attacks can be carried out by any FL par-
ticipant. Due to the distributed nature of FL, data poisoning
attacks are generally less effective than model poisoning
attacks [26], [27], [28], [29]. The impact on the FL model
depends on the extent to which the backdoor participants
engage in the attacks, and the amount of training data being
poisoned. A recent work shows that poisoning edge-case
(low probability) training samples are more effective [24].