Amplification attacks: other people's servers, your victim's link

A reflection attack sends requests to servers on the internet with the victim's address forged as the source, so the servers send their replies to the victim. An amplification attack picks services whose replies are much larger than the requests, so every byte the attacker sends turns into many bytes arriving at the victim. Together they let a modest botnet fill very large links.

Section Attack types Updated Needs spoofing + open services

Quick facts

Reflection
Replies are sent to a forged source address: the victim
Amplification
Each reply is larger than the request that caused it
Protocol
Almost always UDP, which needs no handshake
Typical services abused
Time servers, memcached, DNS resolvers and other UDP services left open
Also searched as
amplified DNS attack, reflector attack
Root causes
Networks that allow spoofed sources (BCP 38) and services that answer anyone

The general problem is described in RFC 4732; individual vectors have their own pages below.

How a reflected, amplified DDoS attack works

  1. The attacker finds open services. Servers that answer UDP queries from anyone on the internet.
  2. It sends them small queries with a forged source. The source is the victim's address; see IP spoofing.
  3. The services reply to the victim. Each reply is many times the size of the query.
  4. Thousands of services reply at once. The victim's link fills with traffic from servers that look entirely legitimate.

The victim sees traffic from real, reachable servers, not from the attacker. Blocking those servers one by one is slow, and some of them may be services the victim also needs.

WHAT REFLECTION ADDS TO A FLOOD REFLECTED AMPLIFIED SOURCES LOOK LEGITIMATE NTP MEMCACHED DNS PLAIN UDP FLOOD // the last row is the attacker's own traffic, sent directly
// three reflection vectors compared with a direct flood

The amplification factor

The amplification factor is the ratio between the size of the reply and the size of the request that triggered it. A factor of 10 means each byte the attacker sends with a forged source turns into ten bytes at the victim. For exposed memcached servers, GitHub reported a factor of up to 51,000.

The factor decides how much bandwidth an attacker needs. With a large factor, a modest connection and a list of reflectors can produce a flood far larger than the attacker's own link, which is why protocols with large factors are found and abused quickly once they are exposed.

Reflection without amplification

Reflection and amplification are often named together, and they are two different properties. Reflection hides the source: the victim receives traffic from the reflecting servers, not from the attacker. Amplification multiplies the volume. An attack can reflect without amplifying, for example by bouncing TCP SYN packets off ordinary servers, which then send their SYN-ACK replies to the victim.

Reflection alone gives the attacker cover but no extra bandwidth, so most large reflected attacks pick protocols that do both. Either way, both depend on forged source addresses, which is why source address validation stops them at the origin. The oldest named case, which used a whole network as the amplifier, is explained in what is a smurf attack.

Common amplification vectors

  • NTP. Time servers that answer status queries with long lists; see NTP amplification.
  • Memcached. Caching servers exposed to the internet over UDP; behind the 1.35 Tbps attack on GitHub in 2018 (GitHub). See memcached DDoS.
  • DNS. Open resolvers answering large queries; DNS amplification is covered on Bridge DNS, the naming layer of the network.
  • Other UDP services. Device discovery, directory and chat protocols left reachable from the internet.

Amplified DNS attack: open resolvers

An amplified DNS attack uses DNS servers as the amplifiers: a short query carrying the victim's address as its source brings back a much longer answer, which goes to the victim. RFC 5358 points to open recursive nameservers, which answer any address, and says that with EDNS0 the amplification factor could be around 80. That is a 2008 figure for that setup, not for every DNS server today.

Authoritative servers can be abused too, but in 2008 RFC 5358 judged their amplification potential greatly reduced.

Reflector attack: the server in the middle

In a reflector attack, the reflector is the third-party server that bounces the traffic. It gets a request with a forged source address, the victim's, and replies to the victim, so the victim sees the reflector's address, not the attacker's. RFC 5358 calls the DNS version of this a reflector attack.

What an amplification attack looks like from the receiving end

Traffic arrives from thousands of addresses that all belong to real servers, all sending from the same service port, such as the port of a time or caching service. The packets are replies to questions you never asked, so there is no matching outbound traffic from your side.

Because the packets are large and arrive together, the link fills long before your servers are busy. Graphs from your provider show the inbound side at capacity while CPU and request rates on your servers stay normal or fall, since real requests can no longer get through.

How amplification attacks are stopped

At the victim, only capacity in front of the link helps: the provider or an edge network absorbs or filters the flood before it arrives. Because the traffic is UDP from known service ports, it is comparatively easy to recognise; the difficulty is volume.

Across the internet, the attacks are prevented at their two sources: networks that drop packets with forged sources (BCP 38) and operators who close or restrict services that answer anyone. Every server you take off the open internet is one fewer amplifier.

Making sure your servers are not amplifiers

  1. List UDP services reachable from the internet. Scan your own address ranges from outside.
  2. Close what does not need to be public. Memcached, management protocols and discovery services rarely do.
  3. Restrict the rest. Answer only your own clients, and disable query types that return large lists.
  4. Watch outbound traffic. A sudden rise in replies to many destinations is a sign a service is being used.

To keep an open recursive resolver out of an amplified DNS attack, RFC 5358 recommends offering recursive lookups only to the intended clients: filter by source address, by incoming interface, or by signed queries (TSIG or SIG(0)).

Questions

What is an amplification attack?

An amplification attack sends small requests with a forged source address to services whose replies are much larger, so the replies flood the victim with far more traffic than the attacker sent.

What is the difference between reflection and amplification?

Reflection means replies go to a forged victim address instead of the sender. Amplification means each reply is larger than the request. Most such attacks use both.

How do you stop a reflected DDoS attack?

The victim needs capacity or filtering upstream of its link. Across the internet, the fix is source address validation by networks and closing services that answer anyone.

What is an amplified DNS attack?

An amplified DNS attack sends short DNS queries with the victim's address forged as the source to servers that answer anyone, such as open recursive resolvers. Each answer is longer than its query and goes to the victim.

What is a reflector in a reflector attack?

In a reflector attack, a reflector is a server that receives a request with the victim's forged source address and sends its reply to the victim. It is a real server run by someone else, so the flood comes from legitimate addresses.